원클릭으로 Manus에서 모든 스킬 실행

vercel-cost-guard

스타2

포크0

업데이트2026년 2월 17일 17:01

Audits Next.js/Vercel projects for 22 cost-causing patterns across bandwidth, compute, invocations, and image optimization. Scans for large uncompressed media, missing preload attributes, API routes without caching, AI streaming without timeouts, database N+1 queries, unscoped middleware, accidental dynamic rendering, payload limit violations, missing ISR/static generation, and configuration issues. Use when reviewing a Next.js project for cost issues, preparing for launch, or after receiving an unexpected Vercel bill.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

jshchnz

jshchnz/vercel-cost-guard

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Vercel Cost Guard

Audit a Next.js/Vercel project for patterns that cause high hosting costs — bandwidth, compute, and invocation charges.

Vercel's pay-per-use pricing means code patterns directly affect your hosting bill. A missing preload="none", an unscoped middleware, or a single cookies() call can silently multiply costs. This audit catches these patterns before they show up on an invoice.

How to Run the Audit

Perform the following 22 checks in order using your built-in tools (Glob, Grep, Read). After completing all checks, compile findings into the report format at the bottom.

Report generation rules:

For every CRITICAL finding, include a copy-paste-ready fix (shell command, code snippet, or config change) drawn from reference/optimization-guide.md. Substitute the project's actual file paths into the commands.
Estimate dollar impact where possible. Use the cost formulas from reference/vercel-pricing.md. Assume 50K monthly visitors as a baseline unless the project's traffic is known. State your assumption.
Track positive findings ("GOOD" patterns from Checks 2, 8, and 11) — these are used in the report when the project has no critical or warning issues.

Important: Search .ts and .js files in addition to .tsx and .jsx — blog content and HTML strings are often defined in TypeScript template literals, not just JSX files.

Check 1: Public Directory Assets

Use Glob to find all files in public/. For each media file, note its size.

CRITICAL — Files > 10 MB: Any file over 10 MB in public/ will cause significant bandwidth costs under traffic. Each visitor downloads it fully.

Fix: Compress the file, convert to a smaller format, or move to an external CDN (Cloudflare R2, AWS S3).

WARNING — Files 5-10 MB: Large but not extreme. Should be compressed before any launch or promotion.

CRITICAL — Uncompressed formats: Flag any files with these extensions — they have compressed alternatives that are 2-20x smaller:

.wav, .flac → convert to .mp3 or .m4a
.mov, .avi → convert to .mp4 (H.264)
.bmp, .tiff, .tif, .raw, .psd → convert to .webp or .avif

Calculate the total size of all media files in public/ and list the top 5 largest files.

For detailed compression commands, see reference/optimization-guide.md.

Check 2: Media Preload Attributes

Use Grep to search all .tsx, .jsx, .ts, .js, and .html files (excluding node_modules/ and .next/) for <audio and <video elements.

CRITICAL — preload="auto": Pattern: <(audio|video)[^>]*preload\s*=\s*["']auto["'] The browser downloads the ENTIRE file on page load. This is the most expensive setting.

Fix: Change to preload="none".

WARNING — preload="metadata": Pattern: <(audio|video)[^>]*preload\s*=\s*["']metadata["'] The browser downloads file headers to read duration/codec info. For large files, this can be several MB per element.

Fix: Change to preload="none" for cost safety.

CRITICAL — Missing preload attribute: Find <audio and <video tags that do NOT have a preload attribute. Many browsers default to preload="auto" when the attribute is missing, which downloads the entire file.

Fix: Add preload="none" explicitly.

GOOD — preload="none": Note any elements correctly using preload="none" as positive findings.

Check 3: Next.js Config

Use Glob to find next.config.ts, next.config.mjs, or next.config.js. Read the file.

WARNING — Missing images.minimumCacheTTL (Next.js 14/15): Without this setting, optimized images expire quickly and must be re-optimized on the next request (costs compute + bandwidth). Next.js 16+ increased the default from 60s to 14,400s (4 hours), so this is less critical on v16+ but still important on v14/v15 where the default is only 60s.

Fix: Add minimumCacheTTL: 2592000 (30 days) to the images config.

WARNING — Low cache TTL (< 86400 seconds / 1 day): Even with the setting present, a value under 1 day causes frequent re-optimization.

Fix: Increase to at least 2592000 (30 days) for static images.

INFO — Missing AVIF format: If images.formats does not include 'image/avif', images are 20-50% larger than they could be.

Fix: Add formats: ['image/avif', 'image/webp'] to images config.

WARNING — Missing custom Cache-Control headers: Check if the config has a headers() function with Cache-Control for static assets. Without custom headers, static assets may not be cached effectively at the CDN edge.

Fix: Add long-lived cache headers for static file extensions. See reference/optimization-guide.md.

WARNING — Overly broad remotePatterns: Check the images.remotePatterns array. Flag if any entry uses hostname: '**' or a very permissive wildcard (e.g., hostname: '*.com'). Broad remote patterns allow any external source to trigger image optimization (billed per transformation — see reference/vercel-pricing.md). Wildcard hostnames also create SSRF/DoS risk against the image optimizer (CVE-2025-59471).

Fix: Restrict remotePatterns to specific, known hostnames your app actually uses.

Check 4: Image Optimization

WARNING — Raw <img> tags instead of next/image: Use Grep to find <img tags in app/, components/, and pages/ directories (excluding node_modules/). The Next.js <Image> component provides automatic optimization, lazy loading, and modern format conversion.

Fix: Replace <img> with <Image> from next/image.

WARNING — GIF files with Image component: Grep for Image component usage with .gif sources. Next.js does NOT optimize GIFs — they pass through at full size.

Fix: Convert GIFs to .mp4/.webm video, or use a static image.

INFO — Missing quality prop on Image components: Grep for <Image usage without a quality prop. The default is 75, which is usually fine, but explicitly setting it to 75 or lower can reduce sizes further.

WARNING — Large PNGs in public/: Use Glob to find .png files > 500 KB in public/. PNGs are lossless and often 2-4x larger than WebP/AVIF.

Fix: Convert to WebP (25-34% smaller) or AVIF (up to 50% smaller).

Check 5: API Route Cache Headers

Find API route files in app/api/ or pages/api/.

WARNING — GET routes missing Cache-Control: Grep for files that export a GET handler but don't contain Cache-Control. Every uncached GET request invokes the serverless function, costing compute time and bandwidth.

Fix: Add Cache-Control: public, s-maxage=60, stale-while-revalidate=300 (adjust TTL to your needs).

Check 6: Anti-Patterns

WARNING — Server components calling own API routes: Grep for fetch('/api/ or fetch("/api/ in files that do NOT contain 'use client' or "use client". Server components calling their own API routes cause double invocations (the server component invocation + the API route invocation).

Fix: Import the API logic directly instead of fetching via HTTP.

INFO — Excessive Link prefetching: Grep for <Link with prefetch={true} or prefetch=\{true\}. In Next.js 15+, prefetch={true} is deprecated — the default behavior ("auto") already prefetches visible links in the viewport. Explicit prefetch={true} is redundant and may increase bandwidth by fully prefetching dynamic routes.

Fix: Remove prefetch={true} (use the default). Use prefetch={false} for infrequently visited pages.

INFO — Client-side fetching without caching library: Check if package.json includes swr or @tanstack/react-query. If neither is present and client components use fetch(), data may be re-fetched unnecessarily on every render.

Fix: Add SWR or React Query for automatic caching and request deduplication.

Check 7: Vercel Spend Management

This check is always included in the report, regardless of other findings.

Vercel's Spend Management sets an on-demand budget with alerts at 50%, 75%, and 100%. It is enabled by default on Pro plans. Check the current default limit at the Spend Management docs — do not assume a specific dollar amount, as defaults change.

Remind the user to:

Go to Vercel Dashboard → Settings → Billing → Spend Management
Review the current limit and lower it if appropriate (e.g., $20-50 for hobby projects)
Consider enabling "Pause production deployment" at 100% for a hard spending cap
Configure alert notifications for usage spikes

For pricing details, see reference/vercel-pricing.md.

Check 8: Middleware Scope

Use Glob to find middleware.ts, middleware.js, proxy.ts, or proxy.js at the project root (or src/). Next.js 16 renamed middleware.ts to proxy.ts — check for both.

If no middleware/proxy file exists, skip this check.

If one exists, Read the file and check for a matcher config.

WARNING — No matcher config: Without a matcher, middleware runs on every request — including _next/static, _next/image, and other asset requests. This can multiply edge invocations by 10x or more.

Fix: Add a config.matcher that excludes static assets. See reference/optimization-guide.md.

WARNING — matcher does not exclude _next/static or _next/image: Even with a matcher, if it doesn't exclude Next.js internal asset paths, middleware still runs on those requests unnecessarily.

Fix: Use a negative lookahead pattern: '/((?!_next/static|_next/image|favicon.ico).*)'

GOOD — Properly scoped middleware: Note if middleware has a well-scoped matcher that excludes static assets.

Check 9: Accidental Dynamic Rendering

Use Grep to search page.tsx and page.ts files in the app/ directory for patterns that force dynamic rendering.

Note: In Next.js 15+, cookies(), headers(), and searchParams are async and must be awaited. Grep patterns should catch both sync and async usage.

WARNING — cookies() usage from next/headers: Pattern: (await\s+)?cookies in files that import from next/headers A single cookies() call forces the entire page to render on every request instead of being statically cached.

Fix: Move cookie-dependent logic to a client component, or handle in middleware.

WARNING — headers() usage from next/headers: Pattern: (await\s+)?headers in files that import from next/headers Same dynamic opt-in as cookies().

Fix: Extract needed header values in middleware and pass via rewrites/cookies.

WARNING — searchParams in page component signature: Pattern: searchParams in the page component's props type or destructuring (in Next.js 15+, this is a Promise that must be awaited) Accessing searchParams in a server component forces dynamic rendering for every unique URL.

Fix: Use generateStaticParams for known parameter values, or move filtering logic to a client component.

INFO — connection() or unstable_noStore() usage: Also check for connection() from next/server (Next.js 15+) or the legacy unstable_noStore() — both explicitly opt out of static rendering.

Fix: Use revalidate instead for periodic freshness where possible.

For detailed before/after examples, see reference/optimization-guide.md.

Check 10: Sequential Async in Serverless

Use Grep to search API route files (app/api/**/route.ts, app/api/**/route.js) and server action files for functions with 3 or more sequential await statements.

INFO — 3+ sequential await in a single function: Each sequential await keeps the serverless function running (and billing) while waiting on I/O. Three sequential 200ms operations take 600ms; parallelized they take 200ms — a 3x reduction in billed duration.

Fix: Wrap independent operations in Promise.all(). See reference/optimization-guide.md.

Note: Only flag await statements that appear to be independent (not where one result feeds into the next). Use judgment — const user = await getUser(); const posts = await getPostsByUser(user.id); is necessarily sequential.

INFO — Missing maxDuration route segment config: If no export const maxDuration = N is set on API routes or pages with Server Actions, functions run up to the platform default (which can be long on Pro plans). Adding maxDuration caps execution time and prevents runaway billing from slow upstream services.

Fix: Add export const maxDuration = 10 (or appropriate value) to API routes and pages with Server Actions.

Check 11: Missing ISR / generateStaticParams

Use Glob to find dynamic route directories in app/ — directories containing [param], [...param], or [[...param]] segments.

For each dynamic route, Read the page.tsx/page.ts file and check for:

WARNING — Dynamic route without generateStaticParams or revalidate: If a dynamic route page has neither export async function generateStaticParams nor export const revalidate, every visit triggers a serverless function invocation.

Fix: Add generateStaticParams to pre-render known pages at build time, or add export const revalidate = 3600 for ISR.

GOOD — Has generateStaticParams: Note pages that correctly pre-render dynamic routes.

GOOD — Has revalidate export: Note pages using ISR with a reasonable revalidation interval.

Check 12: AI Agent & Streaming Response Patterns

Use Grep to search API routes (app/api/**/route.ts, app/api/**/route.js) and server action files for AI API usage patterns.

CRITICAL — Long-running AI streaming without timeout caps: Grep for AI provider imports: openai, anthropic, @anthropic-ai/sdk, @ai-sdk, keywords like chat/completions, stream: true, messages.create.

For any file using AI APIs, Read the file and check:

Is there an export const maxDuration = N statement?
If streaming (stream: true), is there an AbortSignal or timeout control?

Without maxDuration, functions can run up to 300s on Pro (costs add up fast). Without timeout caps, long-running AI requests can cause bills to spike from $20 to $1000+/month.

Fix: Add export const maxDuration = 30 (or appropriate value based on model). Add AbortSignal with timeout:

export const maxDuration = 30;

export async function POST(req: Request) {
  const controller = new AbortController();
  const timeoutId = setTimeout(() => controller.abort(), 25000);

  try {
    const completion = await openai.chat.completions.create(
      { model: 'gpt-4', messages: [...], stream: true },
      { signal: controller.signal }
    );
    clearTimeout(timeoutId);
    return new StreamingTextResponse(OpenAIStream(completion));
  } catch (err) {
    clearTimeout(timeoutId);
    if (err.name === 'AbortError') {
      return Response.json({ error: 'Request timeout' }, { status: 504 });
    }
    throw err;
  }
}

WARNING — Streaming responses without proper chunking: Check if files with stream: true return a ReadableStream or StreamingTextResponse. If they collect chunks into an array or string before returning, they're buffering in memory (defeats streaming benefits and can cause memory issues).

Fix: Use StreamingTextResponse from ai package or return a proper ReadableStream.

WARNING — AI responses in middleware: Grep for AI API calls (openai, anthropic, AI-related imports) in middleware.ts, middleware.js, proxy.ts, or proxy.js.

Middleware runs on Edge runtime with a 30s timeout limit. AI calls often take 10-30+ seconds and are unpredictable. This causes frequent 504 timeouts.

Fix: Move AI logic to API routes or Server Actions. Use middleware only for lightweight checks (e.g., cache lookups).

WARNING — Missing maxDuration on AI routes: For any route with AI API calls that doesn't have export const maxDuration, flag it. On Pro plans, the default is 300s. On Hobby, it's 60s. Without an explicit cap, you're vulnerable to runaway billing.

Fix: Add export const maxDuration = N where N is appropriate for the AI model being used (GPT-3.5: 15s, GPT-4: 30s, Claude Opus: 60s).

For detailed AI cost optimization patterns, see reference/ai-costs.md.

Check 13: Request/Response Payload Limits

Vercel has a 4.5 MB limit on request and response payloads for standard functions. Exceeding this causes FUNCTION_PAYLOAD_TOO_LARGE errors (413).

CRITICAL — Large JSON payloads in API routes: Use Grep to find API routes that accept file uploads or large data:

Pattern: await req.json(), await req.formData(), file upload patterns

For routes that accept uploads or large POST bodies, check if they handle files > 4.5 MB or use Base64 encoding (which inflates size by 33%).

Calculate potential payload size:

Image files: 3.5 MB image → 4.67 MB as Base64 (exceeds limit!)
Form data with multiple files
Large JSON arrays
Fix: Use streaming functions for payloads > 4.5 MB:

export async function POST(req: Request) {
  // Don't call req.json() — work with raw stream
  const stream = req.body;

  await uploadToStorage(stream); // Stream directly to S3/R2
  return Response.json({ success: true });
}

WARNING — Base64 file encoding in API routes: Grep for Base64 encoding patterns: btoa(, atob(, base64, data:image, Buffer.from(.*base64).

Base64 increases file size by ~33%. A 3.5 MB file becomes 4.67 MB (exceeds limit).

Fix: Use FormData with binary files instead of Base64, or stream directly to external storage from the client:

// Bad: Base64 (inflates size)
const base64 = await fileToBase64(image);
fetch('/api/upload', { body: JSON.stringify({ image: base64 }) });

// Good: FormData (raw binary)
const formData = new FormData();
formData.append('image', file);
fetch('/api/upload', { body: formData });

// Better: Direct upload to S3/R2 (bypasses Vercel entirely)
const { uploadUrl } = await fetch('/api/upload-url').then(r => r.json());
await fetch(uploadUrl, { method: 'PUT', body: file });

INFO — Missing streaming function detection: Check if the project has any routes using streaming for large responses. If not, and there are API routes returning large data, suggest considering streaming.

Fix: For responses > 4.5 MB, use streaming:

export async function GET() {
  const largeData = await fetchLargeDataset();

  const stream = new ReadableStream({
    start(controller) {
      controller.enqueue(JSON.stringify(largeData));
      controller.close();
    },
  });

  return new Response(stream, {
    headers: { 'Content-Type': 'application/json' },
  });
}

Check 14: Database N+1 Query Patterns

Use Grep to search for common N+1 query patterns in API routes, Server Actions, and server components.

WARNING — Loop with await inside: Pattern: for.*await, \.map.*async.*=>.*await, .forEach.*await

These patterns often indicate sequential database queries inside a loop — each iteration waits for a query result before starting the next.

Example of bad pattern:

for (const post of posts) {
  post.author = await db.findUser(post.authorId); // N queries
}

Fix: Use Promise.all() for independent queries, or use a single query with JOIN/IN clause:

// Option 1: Promise.all for independent queries
const authors = await Promise.all(
  posts.map(post => db.findUser(post.authorId))
);

// Option 2: Single query with IN clause (better)
const authorIds = [...new Set(posts.map(p => p.authorId))];
const authors = await db.findMany({ where: { id: { in: authorIds } } });

WARNING — Multiple sequential ORM find() calls: Grep for multiple sequential database query calls in the same function:

Prisma: multiple prisma.*.findMany() or prisma.*.findUnique() in sequence
Drizzle: multiple db.select().from() in sequence
Supabase: multiple supabase.from().select() in sequence

Each query is a network round-trip (20-50ms same-region, 100-200ms cross-region). Sequential queries multiply function duration.

Fix: Use ORM relation loading or parallel queries:

// Bad: Sequential (300ms total if 3 × 100ms per query)
const user = await prisma.user.findUnique({ where: { id } });
const posts = await prisma.post.findMany({ where: { authorId: id } });
const comments = await prisma.comment.findMany({ where: { authorId: id } });

// Good: Parallel (100ms total, all run concurrently)
const [user, posts, comments] = await Promise.all([
  prisma.user.findUnique({ where: { id } }),
  prisma.post.findMany({ where: { authorId: id } }),
  prisma.comment.findMany({ where: { authorId: id } }),
]);

// Better: Single query with relations
const user = await prisma.user.findUnique({
  where: { id },
  include: { posts: true, comments: true },
});

INFO — Missing connection pooling: Check package.json for database client packages: @prisma/client, drizzle-orm, pg, postgres, mysql2, @supabase/supabase-js.

Then check:

If using Prisma: Look for @prisma/extension-accelerate (Prisma Accelerate provides connection pooling)
If using Postgres: Check for PgBouncer or Supabase Pooler config in env vars
If using PlanetScale/Neon: Pooling is built-in (note this as a positive finding)

Without connection pooling, every function invocation creates a new database connection (100-500ms overhead on cold starts).

Fix: Enable connection pooling:
- Prisma: Add Prisma Accelerate
- Supabase: Enable Pooler in dashboard and use pooler connection string
- Self-hosted Postgres: Add PgBouncer
- PlanetScale/Neon: No action needed (built-in)

For detailed database optimization patterns, see reference/database-optimization.md.

Check 15: Missing Compression Headers

Vercel automatically compresses responses > 1 KB, but explicit configuration ensures optimal compression for all content types.

WARNING — API routes with large JSON responses lacking compression: Find API routes (app/api/**/route.ts) that return large datasets (arrays, objects).

While Vercel auto-compresses JSON > 1 KB, responses near the 4.5 MB limit should explicitly hint compression. Gzip/Brotli can reduce JSON by 70-90%.

Fix: Add explicit content headers (Vercel handles the actual compression):

export async function GET() {
  const largeData = await getLargeDataset();

  return new Response(JSON.stringify(largeData), {
    headers: {
      'Content-Type': 'application/json',
      'Cache-Control': 'public, s-maxage=60',
    },
  });
}

WARNING — Middleware doubling Fast Data Transfer: Check if middleware.ts/proxy.ts reads or modifies request/response bodies (pattern: await req.json(), await req.text(), NextResponse.json(modifiedBody)).

When middleware reads/modifies bodies, data is transferred multiple times:

Client → Edge (counted as FDT)
Edge → Middleware (processes body) → Edge (counted again)
Edge → Origin → Edge → Client (counted again)

Result: Same data counted 2-3x in bandwidth billing.

Fix: Don't process bodies in middleware. Use middleware only for headers, redirects, and rewrites:

// Bad: Reading/modifying body in middleware
export async function middleware(req: NextRequest) {
  const body = await req.json(); // Triggers FDT
  const modified = transformBody(body);
  return NextResponse.json(modified); // Triggers FDT again
}

// Good: Only inspect headers
export async function middleware(req: NextRequest) {
  const authHeader = req.headers.get('authorization');
  if (!authHeader) {
    return new Response('Unauthorized', { status: 401 });
  }
  return NextResponse.next();
}

INFO — SVG files without compression: Use Glob to find .svg files in public/. SVG files are XML-based text and compress 60-80% with gzip, but are often served uncompressed.

Fix: Ensure next.config.js has compression headers for SVG mime type, or pre-compress SVGs at build time:

// next.config.js
async headers() {
  return [
    {
      source: '/:all*(svg)',
      headers: [
        {
          key: 'Content-Type',
          value: 'image/svg+xml',
        },
        {
          key: 'Cache-Control',
          value: 'public, max-age=31536000, immutable',
        },
      ],
    },
  ];
}

Check 16: ISR Read/Write Cost Patterns

Use Grep to find pages with ISR configuration (export const revalidate).

WARNING — ISR with very short revalidate (< 60s): Grep for export const revalidate\s*=\s*\d+ and extract the value.

If revalidate is less than 60 seconds, the page revalidates very frequently. Each revalidation triggers:

1 function invocation
Compute time to render the page
Fast Origin Transfer (page HTML from origin to edge)

For a page revalidating every 30 seconds under constant traffic:

Revalidations per hour: 120
Revalidations per day: 2,880
Revalidations per month: ~86,400 function invocations

Fix: Increase to at least 300 seconds (5 minutes) unless real-time freshness is critical:

// Bad: Revalidates every 30 seconds
export const revalidate = 30;

// Good: Revalidates every 5 minutes
export const revalidate = 300;

// Better: Revalidates every hour
export const revalidate = 3600;

INFO — ISR without on-demand revalidation: For pages with export const revalidate, check if the project has any usage of revalidatePath() or revalidateTag() (Grep for these patterns in API routes or Server Actions).

If not found, the project is only using time-based revalidation. This means content can't be updated immediately when changed — you have to wait until the revalidation period expires.

Fix: Add on-demand revalidation for time-sensitive content updates:

// app/api/revalidate/route.ts
import { revalidatePath } from 'next/cache';

export async function POST(req: Request) {
  const { path } = await req.json();

  // Trigger immediate revalidation when content changes
  revalidatePath(path);

  return Response.json({ revalidated: true, now: Date.now() });
}

// Hybrid approach: time-based fallback + on-demand for immediate updates
export const revalidate = 3600; // Fallback every hour

// In page.tsx or CMS webhook: call revalidatePath when content changes

Check 17: Vercel Fluid Compute Detection

This check reminds the user to enable Vercel Fluid Compute if not already enabled.

INFO — Fluid Compute eligibility check: Check if the project is deployed on Vercel (look for vercel.json, .vercel/ directory, or VERCEL in environment variables).

If the project is on Vercel and not on the Hobby plan, Fluid Compute can reduce function costs by 20-30% for I/O-heavy workloads (database queries, API calls, AI workloads).

Fluid Compute uses Active CPU pricing:

Active CPU time billed at standard rate
I/O wait time (database queries, API calls) billed at lower rate

For functions with database queries or AI API calls, most of the time is I/O wait, making Fluid Compute very cost-effective.

Fix: Enable Fluid Compute in the Vercel Dashboard:
1. Go to Dashboard → [Project] → Settings → Functions
2. Scroll to "Function Duration and Fluid Compute"
3. Toggle "Enable Fluid Compute" to ON
4. Deploy the project (applies to new deployments)

INFO — I/O-heavy functions without optimization: For functions with database queries or external API calls, remind the user that Fluid Compute bills active CPU at a higher rate than I/O wait time. Parallelizing I/O operations reduces total function time (and thus total cost).

Fix: Use Promise.all() to parallelize independent I/O operations (see Check 10 and Check 14).

Check 18: Function Timeout Patterns

Use Grep to find routes with long-running operations that may need timeout configuration.

WARNING — AI/streaming routes without maxDuration on Hobby: For routes with AI API calls (detected in Check 12), check if there's an export const maxDuration statement.

If not, and the project appears to be on the Hobby plan (no Pro plan indicators like vercel.json with custom regions or enterprise features), warn that the Hobby plan has a 60s hard timeout.

AI models like GPT-4 streaming can take 30-60 seconds. Without explicit timeout configuration, requests may hit the 60s limit and fail with 504 errors.

Fix: Add export const maxDuration = 60 (or lower) for Hobby plan, or upgrade to Pro for longer timeouts:

// For Hobby plan (60s max)
export const maxDuration = 60;

// For Pro plan (300s default, up to 900s max)
export const maxDuration = 300;

WARNING — Database operations without timeout: For routes with database queries, check if the database client has timeout configuration.

Long-running queries (e.g., full table scans, complex joins on large datasets) can max out function duration and cause high costs.

Fix: Add query timeout at the client level:

// Prisma
const prisma = new PrismaClient({
  datasources: {
    db: {
      url: process.env.DATABASE_URL,
    },
  },
  // Add query timeout
  log: [{ level: 'query', emit: 'event' }],
});

// Postgres (pg package)
const client = new Client({
  connectionString: process.env.DATABASE_URL,
  statement_timeout: 10000, // 10 seconds
});

// Supabase
const { data, error } = await supabase
  .from('posts')
  .select('*')
  .abortSignal(AbortSignal.timeout(10000)); // 10 seconds

Check 19: Fetch Deduplication

Use Grep to search for fetch() calls in Server Components (files without 'use client' or "use client").

INFO — Multiple identical fetch() calls: Check for the same URL fetched multiple times in server components or API routes.

Next.js automatically deduplicates fetch() requests with the same URL within a single render, but explicit deduplication makes intent clearer and works across renders.

Fix: Use React.cache() for explicit deduplication:

import { cache } from 'react';

const getUser = cache(async (id: string) => {
  const res = await fetch(`https://api.example.com/users/${id}`);
  return res.json();
});

// Multiple calls to getUser(id) with the same id will only fetch once

INFO — fetch() with cache: 'no-store' in server components: Pattern: fetch(.*cache.*no-store) or fetch(.*cache.*no-cache)

Server components with cache: 'no-store' opt out of automatic fetch deduplication and caching. This causes the same data to be fetched multiple times.

Fix: Remove cache: 'no-store' unless the data is truly per-request:

// Bad: Opts out of caching
const res = await fetch('https://api.example.com/data', {
  cache: 'no-store',
});

// Good: Use default caching or explicit revalidation
const res = await fetch('https://api.example.com/data', {
  next: { revalidate: 60 }, // Cache for 60 seconds
});

// Or remove options entirely (Next.js caches by default)
const res = await fetch('https://api.example.com/data');

Check 20: Build-Time Optimization

Use Read to check package.json for dependencies and next.config.js/next.config.ts for output mode.

INFO — Large dependencies in package.json: Check for heavy dependencies that have lightweight alternatives:

moment → Replace with date-fns (92% smaller)
lodash (without tree-shaking) → Replace with lodash-es or individual imports
Large UI libraries → Check for tree-shaking configuration

These dependencies increase:

Build time (slower deployments)
Function bundle size (slower cold starts)
Overall deployment size
Fix: Replace with smaller alternatives:

# Replace moment with date-fns
npm uninstall moment
npm install date-fns

# Use tree-shakeable lodash
npm uninstall lodash
npm install lodash-es

INFO — Missing output: 'standalone' in next.config: Check if next.config.js/next.config.ts has output: 'standalone'.

Without standalone mode, the deployment bundle includes all of node_modules, which can be very large. Standalone mode only includes production dependencies actually used by the app (~80% size reduction).

Fix: Add output: 'standalone' to next.config.js:

// next.config.js
const nextConfig = {
  output: 'standalone', // Reduces deployment size by ~80%
};

Note: This is automatically enabled for Vercel deployments, but worth checking for self-hosted or custom deployment setups.

Check 21: Connection Pooling

Use Read to check database client configuration (already partially covered in Check 14, but this check focuses on configuration details).

WARNING — Database client without pooling config: For projects using database clients (Prisma, Drizzle, Postgres.js, MySQL2), check if connection pooling is configured.

Grep environment variables (.env.local, .env) for pooling-related config:

Prisma Accelerate: PRISMA_ACCELERATE_URL or @prisma/extension-accelerate in code
Supabase Pooler: Connection string with :6543 port or ?pgbouncer=true
PgBouncer: Separate pooler URL or config
PlanetScale/Neon: Built-in (no config needed, note as positive finding)

Without pooling, each function instance creates a new database connection:

Cold start overhead: 100-500ms
Connection exhaustion: Databases have max connection limits (Postgres default: 100)
Fix: Enable connection pooling:

Prisma:

import { PrismaClient } from '@prisma/client';
import { withAccelerate } from '@prisma/extension-accelerate';

const prisma = new PrismaClient().$extends(withAccelerate());

Supabase:

# .env.local
# Use pooler connection string instead of direct connection
DATABASE_URL=postgresql://user:pass@pooler.supabase.com:6543/postgres?pgbouncer=true

PgBouncer (self-hosted): Set up PgBouncer as a separate service and point your app to it instead of Postgres directly.

INFO — Redis client without connection reuse: Check if the project uses Redis (@upstash/redis, ioredis, redis).

If the Redis client is created inside function handlers (instead of module scope), it's created fresh on every invocation.

Fix: Move Redis client to module scope with singleton pattern:

// Bad: Creates new client on every invocation
export async function GET() {
  const redis = new Redis({ url: process.env.REDIS_URL });
  const data = await redis.get('key');
  return Response.json(data);
}

// Good: Client created once and reused
import { Redis } from '@upstash/redis';

const redis = new Redis({ url: process.env.REDIS_URL }); // Module scope

export async function GET() {
  const data = await redis.get('key');
  return Response.json(data);
}

Check 22: App Router Performance

Use Grep to analyze App Router vs Pages Router patterns and detect performance anti-patterns.

INFO — App Router with many client components: Count files with 'use client' or "use client" in the app/ directory vs total component files.

If the ratio is high (e.g., > 50% of components are client components), the project may not be leveraging App Router benefits (server-first rendering, streaming).

Fix: Move state to fewer client components and use server components for data fetching:

// Bad: Everything is a client component
'use client';

export default function Page() {
  const [data, setData] = useState(null);

  useEffect(() => {
    fetch('/api/data').then(r => r.json()).then(setData);
  }, []);

  return <div>{data?.title}</div>;
}

// Good: Server component for data, client component only for interactivity
// app/page.tsx (Server Component)
async function getData() {
  const res = await fetch('https://api.example.com/data');
  return res.json();
}

export default async function Page() {
  const data = await getData();
  return <ClientComponent data={data} />;
}

// components/ClientComponent.tsx (only interactive parts)
'use client';

export function ClientComponent({ data }) {
  const [isExpanded, setIsExpanded] = useState(false);
  return (
    <div onClick={() => setIsExpanded(!isExpanded)}>
      {data.title}
      {isExpanded && <Details />}
    </div>
  );
}

INFO — Parallel routes or intercepting routes usage: Grep for parallel routes (directories with @ prefix like @modal, @sidebar) or intercepting routes (directories with (..) like (.)post).

These are advanced App Router features that can add complexity. If used, ensure proper layout optimization and loading states to avoid performance issues.

Fix: Ensure proper loading states and suspense boundaries:

// app/@modal/loading.tsx
export default function Loading() {
  return <ModalSkeleton />;
}

// app/layout.tsx
export default function Layout({ children, modal }) {
  return (
    <>
      {children}
      <Suspense fallback={<ModalSkeleton />}>
        {modal}
      </Suspense>
    </>
  );
}

If parallel/intercepting routes are not adding clear value, consider simplifying to standard routes.

Report Format

After completing all 22 checks, compile findings using the appropriate template below.

If there are any CRITICAL or WARNING findings, use Template A. If there are 0 critical AND 0 warnings, use Template B.

Template A: Standard Report

# Vercel Cost Guard

**[project name]** — [today's date] — [N] critical · [N] warning · [N] info

---

## Critical Issues

[For each CRITICAL finding, render a detail block like the one below.
Order by estimated savings descending — highest-impact issue first.]

### [CHECK_NAME]: [short description]

**File:** `[path/to/file:line]`

**Impact:** [1-2 sentence explanation of why this is expensive, with dollar estimate.
Use formulas from reference/vercel-pricing.md. State the traffic assumption, e.g.
"At 50K monthly visitors, this 52 MB file transfers ~2.6 TB/mo (~$240 in overage)."]

**Fix:**
```[language or bash]
[Exact command or code snippet to fix the issue.
Use the project's actual file paths.
Pull specific commands from reference/optimization-guide.md.]

Estimated savings: ~$[amount]/mo (or ~[X]% bandwidth reduction)

[Repeat for each critical issue, separated by ---]

Warnings

[Render ALL warning findings in a compact table. One row per finding. Do not include full fix details — just a short actionable summary. Order by estimated impact descending.]

#	Check	File	Issue	Fix
1	[CHECK_NAME]	`[path]`	[1-line description]	[1-line action]
2	...	...	...	...

Info

[If there are 0 info findings, omit this section entirely. Otherwise, render as a compact bullet list.]

[CHECK_NAME] — [path]: [description and suggested action]
...

Implementation Checklist

[Consolidate every critical and warning finding into a single numbered checklist. Critical items first, then warnings. Each item gets a copy-paste command or code snippet. Group related items if sensible (e.g., multiple files needing the same conversion).]

1. [Action description — e.g., "Convert hero.mov to MP4"] Saves ~$[amount]/mo
```
[exact command or code change with real file paths]
```
2. [Next action] Saves ~$[amount]/mo
```
[exact command or code change]
```

[Continue for all critical + warning items...]

Cost Optimization Summary

Total estimated monthly savings: ~$[sum of all critical + warning fixes]

Breakdown:

Bandwidth reduction: ~$[amount from media compression, payload optimization]
Compute reduction: ~$[amount from parallelization, timeout caps, ISR]
Invocation reduction: ~$[amount from middleware scoping, static generation, caching]
Image optimization: ~$[amount from image config improvements]

Enable Vercel Fluid Compute in Dashboard → Project Settings → Functions for automatic 20-30% additional savings on function costs (Pro/Enterprise plans only).

Spend Management

Review your Spend Management limit at: Dashboard → Settings → Billing → Spend Management Alerts trigger at 50%, 75%, and 100% of your budget. Enable "Pause production deployment" for a hard cap. Recommended: $20-50 for hobby projects, expected traffic + 50% buffer for production.


---

### Template B: Clean Project

Vercel Cost Guard

[project name] — [today's date] — 0 critical · 0 warnings · [N] info

No critical or warning-level cost issues found.

What You're Doing Right

[List positive findings verified during the audit. Only include items that were actually checked and confirmed — do not fabricate positive findings.]

[checkmark] Media files use compressed formats (mp4, mp3, webp)
[checkmark] All audio/video elements have preload="none"
[checkmark] next.config has minimumCacheTTL set to [value]
[checkmark] Middleware is properly scoped with matcher config
[checkmark] Dynamic routes use generateStaticParams or ISR
[checkmark] API routes include Cache-Control headers
[checkmark] Total public/ media is only ~[size] MB

Optional Improvements

[If there are INFO findings, list them here:]

These won't cause billing issues, but could yield minor improvements:

[CHECK_NAME] — [path]: [description and suggested action]
...

[If there are 0 info findings, replace the list above with:]

No additional recommendations. This project is well-optimized for Vercel's pricing model.

Spend Management

Review your Spend Management limit at: Dashboard → Settings → Billing → Spend Management Alerts trigger at 50%, 75%, and 100% of your budget. Enable "Pause production deployment" for a hard cap. Recommended: $20-50 for hobby projects, expected traffic + 50% buffer for production.

name

vercel-cost-guard

description

Vercel Cost Guard

Audit a Next.js/Vercel project for patterns that cause high hosting costs — bandwidth, compute, and invocation charges.

How to Run the Audit

Perform the following 22 checks in order using your built-in tools (Glob, Grep, Read). After completing all checks, compile findings into the report format at the bottom.

Report generation rules:

For every CRITICAL finding, include a copy-paste-ready fix (shell command, code snippet, or config change) drawn from reference/optimization-guide.md. Substitute the project's actual file paths into the commands.
Estimate dollar impact where possible. Use the cost formulas from reference/vercel-pricing.md. Assume 50K monthly visitors as a baseline unless the project's traffic is known. State your assumption.
Track positive findings ("GOOD" patterns from Checks 2, 8, and 11) — these are used in the report when the project has no critical or warning issues.

Important: Search .ts and .js files in addition to .tsx and .jsx — blog content and HTML strings are often defined in TypeScript template literals, not just JSX files.

Check 1: Public Directory Assets

Use Glob to find all files in public/. For each media file, note its size.

CRITICAL — Files > 10 MB: Any file over 10 MB in public/ will cause significant bandwidth costs under traffic. Each visitor downloads it fully.

Fix: Compress the file, convert to a smaller format, or move to an external CDN (Cloudflare R2, AWS S3).

WARNING — Files 5-10 MB: Large but not extreme. Should be compressed before any launch or promotion.

CRITICAL — Uncompressed formats: Flag any files with these extensions — they have compressed alternatives that are 2-20x smaller:

.wav, .flac → convert to .mp3 or .m4a
.mov, .avi → convert to .mp4 (H.264)
.bmp, .tiff, .tif, .raw, .psd → convert to .webp or .avif

Calculate the total size of all media files in public/ and list the top 5 largest files.

For detailed compression commands, see reference/optimization-guide.md.

Check 2: Media Preload Attributes

Use Grep to search all .tsx, .jsx, .ts, .js, and .html files (excluding node_modules/ and .next/) for <audio and <video elements.

CRITICAL — preload="auto": Pattern: <(audio|video)[^>]*preload\s*=\s*["']auto["'] The browser downloads the ENTIRE file on page load. This is the most expensive setting.

Fix: Change to preload="none".

Fix: Change to preload="none" for cost safety.

Fix: Add preload="none" explicitly.

GOOD — preload="none": Note any elements correctly using preload="none" as positive findings.

Check 3: Next.js Config

Use Glob to find next.config.ts, next.config.mjs, or next.config.js. Read the file.

Fix: Add minimumCacheTTL: 2592000 (30 days) to the images config.

WARNING — Low cache TTL (< 86400 seconds / 1 day): Even with the setting present, a value under 1 day causes frequent re-optimization.

Fix: Increase to at least 2592000 (30 days) for static images.

INFO — Missing AVIF format: If images.formats does not include 'image/avif', images are 20-50% larger than they could be.

Fix: Add formats: ['image/avif', 'image/webp'] to images config.

Fix: Add long-lived cache headers for static file extensions. See reference/optimization-guide.md.

Fix: Restrict remotePatterns to specific, known hostnames your app actually uses.

Check 4: Image Optimization

Fix: Replace <img> with <Image> from next/image.

WARNING — GIF files with Image component: Grep for Image component usage with .gif sources. Next.js does NOT optimize GIFs — they pass through at full size.

Fix: Convert GIFs to .mp4/.webm video, or use a static image.

WARNING — Large PNGs in public/: Use Glob to find .png files > 500 KB in public/. PNGs are lossless and often 2-4x larger than WebP/AVIF.

Fix: Convert to WebP (25-34% smaller) or AVIF (up to 50% smaller).

Check 5: API Route Cache Headers

Find API route files in app/api/ or pages/api/.

Fix: Add Cache-Control: public, s-maxage=60, stale-while-revalidate=300 (adjust TTL to your needs).

Check 6: Anti-Patterns

Fix: Import the API logic directly instead of fetching via HTTP.

Fix: Remove prefetch={true} (use the default). Use prefetch={false} for infrequently visited pages.

Fix: Add SWR or React Query for automatic caching and request deduplication.

Check 7: Vercel Spend Management

This check is always included in the report, regardless of other findings.

Remind the user to:

Go to Vercel Dashboard → Settings → Billing → Spend Management
Review the current limit and lower it if appropriate (e.g., $20-50 for hobby projects)
Consider enabling "Pause production deployment" at 100% for a hard spending cap
Configure alert notifications for usage spikes

For pricing details, see reference/vercel-pricing.md.

Check 8: Middleware Scope

Use Glob to find middleware.ts, middleware.js, proxy.ts, or proxy.js at the project root (or src/). Next.js 16 renamed middleware.ts to proxy.ts — check for both.

If no middleware/proxy file exists, skip this check.

If one exists, Read the file and check for a matcher config.

Fix: Add a config.matcher that excludes static assets. See reference/optimization-guide.md.

Fix: Use a negative lookahead pattern: '/((?!_next/static|_next/image|favicon.ico).*)'

GOOD — Properly scoped middleware: Note if middleware has a well-scoped matcher that excludes static assets.

Check 9: Accidental Dynamic Rendering

Use Grep to search page.tsx and page.ts files in the app/ directory for patterns that force dynamic rendering.

Note: In Next.js 15+, cookies(), headers(), and searchParams are async and must be awaited. Grep patterns should catch both sync and async usage.

Fix: Move cookie-dependent logic to a client component, or handle in middleware.

WARNING — headers() usage from next/headers: Pattern: (await\s+)?headers in files that import from next/headers Same dynamic opt-in as cookies().

Fix: Extract needed header values in middleware and pass via rewrites/cookies.

Fix: Use generateStaticParams for known parameter values, or move filtering logic to a client component.

Fix: Use revalidate instead for periodic freshness where possible.

For detailed before/after examples, see reference/optimization-guide.md.

Check 10: Sequential Async in Serverless

Use Grep to search API route files (app/api/**/route.ts, app/api/**/route.js) and server action files for functions with 3 or more sequential await statements.

Fix: Wrap independent operations in Promise.all(). See reference/optimization-guide.md.

Fix: Add export const maxDuration = 10 (or appropriate value) to API routes and pages with Server Actions.

Check 11: Missing ISR / generateStaticParams

Use Glob to find dynamic route directories in app/ — directories containing [param], [...param], or [[...param]] segments.

For each dynamic route, Read the page.tsx/page.ts file and check for:

Fix: Add generateStaticParams to pre-render known pages at build time, or add export const revalidate = 3600 for ISR.

GOOD — Has generateStaticParams: Note pages that correctly pre-render dynamic routes.

GOOD — Has revalidate export: Note pages using ISR with a reasonable revalidation interval.

Check 12: AI Agent & Streaming Response Patterns

Use Grep to search API routes (app/api/**/route.ts, app/api/**/route.js) and server action files for AI API usage patterns.

For any file using AI APIs, Read the file and check:

Is there an export const maxDuration = N statement?
If streaming (stream: true), is there an AbortSignal or timeout control?

Without maxDuration, functions can run up to 300s on Pro (costs add up fast). Without timeout caps, long-running AI requests can cause bills to spike from $20 to $1000+/month.

Fix: Add export const maxDuration = 30 (or appropriate value based on model). Add AbortSignal with timeout:

export const maxDuration = 30;

export async function POST(req: Request) {
  const controller = new AbortController();
  const timeoutId = setTimeout(() => controller.abort(), 25000);

  try {
    const completion = await openai.chat.completions.create(
      { model: 'gpt-4', messages: [...], stream: true },
      { signal: controller.signal }
    );
    clearTimeout(timeoutId);
    return new StreamingTextResponse(OpenAIStream(completion));
  } catch (err) {
    clearTimeout(timeoutId);
    if (err.name === 'AbortError') {
      return Response.json({ error: 'Request timeout' }, { status: 504 });
    }
    throw err;
  }
}

Fix: Use StreamingTextResponse from ai package or return a proper ReadableStream.

WARNING — AI responses in middleware: Grep for AI API calls (openai, anthropic, AI-related imports) in middleware.ts, middleware.js, proxy.ts, or proxy.js.

Middleware runs on Edge runtime with a 30s timeout limit. AI calls often take 10-30+ seconds and are unpredictable. This causes frequent 504 timeouts.

Fix: Move AI logic to API routes or Server Actions. Use middleware only for lightweight checks (e.g., cache lookups).

Fix: Add export const maxDuration = N where N is appropriate for the AI model being used (GPT-3.5: 15s, GPT-4: 30s, Claude Opus: 60s).

For detailed AI cost optimization patterns, see reference/ai-costs.md.

Check 13: Request/Response Payload Limits

Vercel has a 4.5 MB limit on request and response payloads for standard functions. Exceeding this causes FUNCTION_PAYLOAD_TOO_LARGE errors (413).

CRITICAL — Large JSON payloads in API routes: Use Grep to find API routes that accept file uploads or large data:

Pattern: await req.json(), await req.formData(), file upload patterns

For routes that accept uploads or large POST bodies, check if they handle files > 4.5 MB or use Base64 encoding (which inflates size by 33%).

Calculate potential payload size:

Image files: 3.5 MB image → 4.67 MB as Base64 (exceeds limit!)
Form data with multiple files
Large JSON arrays
Fix: Use streaming functions for payloads > 4.5 MB:

export async function POST(req: Request) {
  // Don't call req.json() — work with raw stream
  const stream = req.body;

  await uploadToStorage(stream); // Stream directly to S3/R2
  return Response.json({ success: true });
}

WARNING — Base64 file encoding in API routes: Grep for Base64 encoding patterns: btoa(, atob(, base64, data:image, Buffer.from(.*base64).

Base64 increases file size by ~33%. A 3.5 MB file becomes 4.67 MB (exceeds limit).

Fix: Use FormData with binary files instead of Base64, or stream directly to external storage from the client:

// Bad: Base64 (inflates size)
const base64 = await fileToBase64(image);
fetch('/api/upload', { body: JSON.stringify({ image: base64 }) });

// Good: FormData (raw binary)
const formData = new FormData();
formData.append('image', file);
fetch('/api/upload', { body: formData });

// Better: Direct upload to S3/R2 (bypasses Vercel entirely)
const { uploadUrl } = await fetch('/api/upload-url').then(r => r.json());
await fetch(uploadUrl, { method: 'PUT', body: file });

Fix: For responses > 4.5 MB, use streaming:

export async function GET() {
  const largeData = await fetchLargeDataset();

  const stream = new ReadableStream({
    start(controller) {
      controller.enqueue(JSON.stringify(largeData));
      controller.close();
    },
  });

  return new Response(stream, {
    headers: { 'Content-Type': 'application/json' },
  });
}

Check 14: Database N+1 Query Patterns

Use Grep to search for common N+1 query patterns in API routes, Server Actions, and server components.

WARNING — Loop with await inside: Pattern: for.*await, \.map.*async.*=>.*await, .forEach.*await

These patterns often indicate sequential database queries inside a loop — each iteration waits for a query result before starting the next.

Example of bad pattern:

for (const post of posts) {
  post.author = await db.findUser(post.authorId); // N queries
}

Fix: Use Promise.all() for independent queries, or use a single query with JOIN/IN clause:

// Option 1: Promise.all for independent queries
const authors = await Promise.all(
  posts.map(post => db.findUser(post.authorId))
);

// Option 2: Single query with IN clause (better)
const authorIds = [...new Set(posts.map(p => p.authorId))];
const authors = await db.findMany({ where: { id: { in: authorIds } } });

WARNING — Multiple sequential ORM find() calls: Grep for multiple sequential database query calls in the same function:

Prisma: multiple prisma.*.findMany() or prisma.*.findUnique() in sequence
Drizzle: multiple db.select().from() in sequence
Supabase: multiple supabase.from().select() in sequence

Each query is a network round-trip (20-50ms same-region, 100-200ms cross-region). Sequential queries multiply function duration.

Fix: Use ORM relation loading or parallel queries:

// Bad: Sequential (300ms total if 3 × 100ms per query)
const user = await prisma.user.findUnique({ where: { id } });
const posts = await prisma.post.findMany({ where: { authorId: id } });
const comments = await prisma.comment.findMany({ where: { authorId: id } });

// Good: Parallel (100ms total, all run concurrently)
const [user, posts, comments] = await Promise.all([
  prisma.user.findUnique({ where: { id } }),
  prisma.post.findMany({ where: { authorId: id } }),
  prisma.comment.findMany({ where: { authorId: id } }),
]);

// Better: Single query with relations
const user = await prisma.user.findUnique({
  where: { id },
  include: { posts: true, comments: true },
});

INFO — Missing connection pooling: Check package.json for database client packages: @prisma/client, drizzle-orm, pg, postgres, mysql2, @supabase/supabase-js.

Then check:

If using Prisma: Look for @prisma/extension-accelerate (Prisma Accelerate provides connection pooling)
If using Postgres: Check for PgBouncer or Supabase Pooler config in env vars
If using PlanetScale/Neon: Pooling is built-in (note this as a positive finding)

Without connection pooling, every function invocation creates a new database connection (100-500ms overhead on cold starts).

Fix: Enable connection pooling:
- Prisma: Add Prisma Accelerate
- Supabase: Enable Pooler in dashboard and use pooler connection string
- Self-hosted Postgres: Add PgBouncer
- PlanetScale/Neon: No action needed (built-in)

For detailed database optimization patterns, see reference/database-optimization.md.

Check 15: Missing Compression Headers

Vercel automatically compresses responses > 1 KB, but explicit configuration ensures optimal compression for all content types.

WARNING — API routes with large JSON responses lacking compression: Find API routes (app/api/**/route.ts) that return large datasets (arrays, objects).

While Vercel auto-compresses JSON > 1 KB, responses near the 4.5 MB limit should explicitly hint compression. Gzip/Brotli can reduce JSON by 70-90%.

Fix: Add explicit content headers (Vercel handles the actual compression):

export async function GET() {
  const largeData = await getLargeDataset();

  return new Response(JSON.stringify(largeData), {
    headers: {
      'Content-Type': 'application/json',
      'Cache-Control': 'public, s-maxage=60',
    },
  });
}

When middleware reads/modifies bodies, data is transferred multiple times:

Client → Edge (counted as FDT)
Edge → Middleware (processes body) → Edge (counted again)
Edge → Origin → Edge → Client (counted again)

Result: Same data counted 2-3x in bandwidth billing.

Fix: Don't process bodies in middleware. Use middleware only for headers, redirects, and rewrites:

// Bad: Reading/modifying body in middleware
export async function middleware(req: NextRequest) {
  const body = await req.json(); // Triggers FDT
  const modified = transformBody(body);
  return NextResponse.json(modified); // Triggers FDT again
}

// Good: Only inspect headers
export async function middleware(req: NextRequest) {
  const authHeader = req.headers.get('authorization');
  if (!authHeader) {
    return new Response('Unauthorized', { status: 401 });
  }
  return NextResponse.next();
}

INFO — SVG files without compression: Use Glob to find .svg files in public/. SVG files are XML-based text and compress 60-80% with gzip, but are often served uncompressed.

Fix: Ensure next.config.js has compression headers for SVG mime type, or pre-compress SVGs at build time:

// next.config.js
async headers() {
  return [
    {
      source: '/:all*(svg)',
      headers: [
        {
          key: 'Content-Type',
          value: 'image/svg+xml',
        },
        {
          key: 'Cache-Control',
          value: 'public, max-age=31536000, immutable',
        },
      ],
    },
  ];
}

Check 16: ISR Read/Write Cost Patterns

Use Grep to find pages with ISR configuration (export const revalidate).

WARNING — ISR with very short revalidate (< 60s): Grep for export const revalidate\s*=\s*\d+ and extract the value.

If revalidate is less than 60 seconds, the page revalidates very frequently. Each revalidation triggers:

1 function invocation
Compute time to render the page
Fast Origin Transfer (page HTML from origin to edge)

For a page revalidating every 30 seconds under constant traffic:

Revalidations per hour: 120
Revalidations per day: 2,880
Revalidations per month: ~86,400 function invocations

Fix: Increase to at least 300 seconds (5 minutes) unless real-time freshness is critical:

// Bad: Revalidates every 30 seconds
export const revalidate = 30;

// Good: Revalidates every 5 minutes
export const revalidate = 300;

// Better: Revalidates every hour
export const revalidate = 3600;

If not found, the project is only using time-based revalidation. This means content can't be updated immediately when changed — you have to wait until the revalidation period expires.

Fix: Add on-demand revalidation for time-sensitive content updates:

// app/api/revalidate/route.ts
import { revalidatePath } from 'next/cache';

export async function POST(req: Request) {
  const { path } = await req.json();

  // Trigger immediate revalidation when content changes
  revalidatePath(path);

  return Response.json({ revalidated: true, now: Date.now() });
}

// Hybrid approach: time-based fallback + on-demand for immediate updates
export const revalidate = 3600; // Fallback every hour

// In page.tsx or CMS webhook: call revalidatePath when content changes

Check 17: Vercel Fluid Compute Detection

This check reminds the user to enable Vercel Fluid Compute if not already enabled.

INFO — Fluid Compute eligibility check: Check if the project is deployed on Vercel (look for vercel.json, .vercel/ directory, or VERCEL in environment variables).

If the project is on Vercel and not on the Hobby plan, Fluid Compute can reduce function costs by 20-30% for I/O-heavy workloads (database queries, API calls, AI workloads).

Fluid Compute uses Active CPU pricing:

Active CPU time billed at standard rate
I/O wait time (database queries, API calls) billed at lower rate

For functions with database queries or AI API calls, most of the time is I/O wait, making Fluid Compute very cost-effective.

Fix: Enable Fluid Compute in the Vercel Dashboard:
1. Go to Dashboard → [Project] → Settings → Functions
2. Scroll to "Function Duration and Fluid Compute"
3. Toggle "Enable Fluid Compute" to ON
4. Deploy the project (applies to new deployments)

Fix: Use Promise.all() to parallelize independent I/O operations (see Check 10 and Check 14).

Check 18: Function Timeout Patterns

Use Grep to find routes with long-running operations that may need timeout configuration.

WARNING — AI/streaming routes without maxDuration on Hobby: For routes with AI API calls (detected in Check 12), check if there's an export const maxDuration statement.

If not, and the project appears to be on the Hobby plan (no Pro plan indicators like vercel.json with custom regions or enterprise features), warn that the Hobby plan has a 60s hard timeout.

AI models like GPT-4 streaming can take 30-60 seconds. Without explicit timeout configuration, requests may hit the 60s limit and fail with 504 errors.

Fix: Add export const maxDuration = 60 (or lower) for Hobby plan, or upgrade to Pro for longer timeouts:

// For Hobby plan (60s max)
export const maxDuration = 60;

// For Pro plan (300s default, up to 900s max)
export const maxDuration = 300;

WARNING — Database operations without timeout: For routes with database queries, check if the database client has timeout configuration.

Long-running queries (e.g., full table scans, complex joins on large datasets) can max out function duration and cause high costs.

Fix: Add query timeout at the client level:

// Prisma
const prisma = new PrismaClient({
  datasources: {
    db: {
      url: process.env.DATABASE_URL,
    },
  },
  // Add query timeout
  log: [{ level: 'query', emit: 'event' }],
});

// Postgres (pg package)
const client = new Client({
  connectionString: process.env.DATABASE_URL,
  statement_timeout: 10000, // 10 seconds
});

// Supabase
const { data, error } = await supabase
  .from('posts')
  .select('*')
  .abortSignal(AbortSignal.timeout(10000)); // 10 seconds

Check 19: Fetch Deduplication

Use Grep to search for fetch() calls in Server Components (files without 'use client' or "use client").

INFO — Multiple identical fetch() calls: Check for the same URL fetched multiple times in server components or API routes.

Next.js automatically deduplicates fetch() requests with the same URL within a single render, but explicit deduplication makes intent clearer and works across renders.

Fix: Use React.cache() for explicit deduplication:

import { cache } from 'react';

const getUser = cache(async (id: string) => {
  const res = await fetch(`https://api.example.com/users/${id}`);
  return res.json();
});

// Multiple calls to getUser(id) with the same id will only fetch once

INFO — fetch() with cache: 'no-store' in server components: Pattern: fetch(.*cache.*no-store) or fetch(.*cache.*no-cache)

Server components with cache: 'no-store' opt out of automatic fetch deduplication and caching. This causes the same data to be fetched multiple times.

Fix: Remove cache: 'no-store' unless the data is truly per-request:

// Bad: Opts out of caching
const res = await fetch('https://api.example.com/data', {
  cache: 'no-store',
});

// Good: Use default caching or explicit revalidation
const res = await fetch('https://api.example.com/data', {
  next: { revalidate: 60 }, // Cache for 60 seconds
});

// Or remove options entirely (Next.js caches by default)
const res = await fetch('https://api.example.com/data');

Check 20: Build-Time Optimization

Use Read to check package.json for dependencies and next.config.js/next.config.ts for output mode.

INFO — Large dependencies in package.json: Check for heavy dependencies that have lightweight alternatives:

moment → Replace with date-fns (92% smaller)
lodash (without tree-shaking) → Replace with lodash-es or individual imports
Large UI libraries → Check for tree-shaking configuration

These dependencies increase:

Build time (slower deployments)
Function bundle size (slower cold starts)
Overall deployment size
Fix: Replace with smaller alternatives:

# Replace moment with date-fns
npm uninstall moment
npm install date-fns

# Use tree-shakeable lodash
npm uninstall lodash
npm install lodash-es

INFO — Missing output: 'standalone' in next.config: Check if next.config.js/next.config.ts has output: 'standalone'.

Fix: Add output: 'standalone' to next.config.js:

// next.config.js
const nextConfig = {
  output: 'standalone', // Reduces deployment size by ~80%
};

Note: This is automatically enabled for Vercel deployments, but worth checking for self-hosted or custom deployment setups.

Check 21: Connection Pooling

Use Read to check database client configuration (already partially covered in Check 14, but this check focuses on configuration details).

WARNING — Database client without pooling config: For projects using database clients (Prisma, Drizzle, Postgres.js, MySQL2), check if connection pooling is configured.

Grep environment variables (.env.local, .env) for pooling-related config:

Prisma Accelerate: PRISMA_ACCELERATE_URL or @prisma/extension-accelerate in code
Supabase Pooler: Connection string with :6543 port or ?pgbouncer=true
PgBouncer: Separate pooler URL or config
PlanetScale/Neon: Built-in (no config needed, note as positive finding)

Without pooling, each function instance creates a new database connection:

Cold start overhead: 100-500ms
Connection exhaustion: Databases have max connection limits (Postgres default: 100)
Fix: Enable connection pooling:

Prisma:

import { PrismaClient } from '@prisma/client';
import { withAccelerate } from '@prisma/extension-accelerate';

const prisma = new PrismaClient().$extends(withAccelerate());

Supabase:

# .env.local
# Use pooler connection string instead of direct connection
DATABASE_URL=postgresql://user:pass@pooler.supabase.com:6543/postgres?pgbouncer=true

PgBouncer (self-hosted): Set up PgBouncer as a separate service and point your app to it instead of Postgres directly.

INFO — Redis client without connection reuse: Check if the project uses Redis (@upstash/redis, ioredis, redis).

If the Redis client is created inside function handlers (instead of module scope), it's created fresh on every invocation.

Fix: Move Redis client to module scope with singleton pattern:

// Bad: Creates new client on every invocation
export async function GET() {
  const redis = new Redis({ url: process.env.REDIS_URL });
  const data = await redis.get('key');
  return Response.json(data);
}

// Good: Client created once and reused
import { Redis } from '@upstash/redis';

const redis = new Redis({ url: process.env.REDIS_URL }); // Module scope

export async function GET() {
  const data = await redis.get('key');
  return Response.json(data);
}

Check 22: App Router Performance

Use Grep to analyze App Router vs Pages Router patterns and detect performance anti-patterns.

INFO — App Router with many client components: Count files with 'use client' or "use client" in the app/ directory vs total component files.

If the ratio is high (e.g., > 50% of components are client components), the project may not be leveraging App Router benefits (server-first rendering, streaming).

Fix: Move state to fewer client components and use server components for data fetching:

// Bad: Everything is a client component
'use client';

export default function Page() {
  const [data, setData] = useState(null);

  useEffect(() => {
    fetch('/api/data').then(r => r.json()).then(setData);
  }, []);

  return <div>{data?.title}</div>;
}

// Good: Server component for data, client component only for interactivity
// app/page.tsx (Server Component)
async function getData() {
  const res = await fetch('https://api.example.com/data');
  return res.json();
}

export default async function Page() {
  const data = await getData();
  return <ClientComponent data={data} />;
}

// components/ClientComponent.tsx (only interactive parts)
'use client';

export function ClientComponent({ data }) {
  const [isExpanded, setIsExpanded] = useState(false);
  return (
    <div onClick={() => setIsExpanded(!isExpanded)}>
      {data.title}
      {isExpanded && <Details />}
    </div>
  );
}

These are advanced App Router features that can add complexity. If used, ensure proper layout optimization and loading states to avoid performance issues.

Fix: Ensure proper loading states and suspense boundaries:

// app/@modal/loading.tsx
export default function Loading() {
  return <ModalSkeleton />;
}

// app/layout.tsx
export default function Layout({ children, modal }) {
  return (
    <>
      {children}
      <Suspense fallback={<ModalSkeleton />}>
        {modal}
      </Suspense>
    </>
  );
}

If parallel/intercepting routes are not adding clear value, consider simplifying to standard routes.

Report Format

After completing all 22 checks, compile findings using the appropriate template below.

If there are any CRITICAL or WARNING findings, use Template A. If there are 0 critical AND 0 warnings, use Template B.

Template A: Standard Report

# Vercel Cost Guard

**[project name]** — [today's date] — [N] critical · [N] warning · [N] info

---

## Critical Issues

[For each CRITICAL finding, render a detail block like the one below.
Order by estimated savings descending — highest-impact issue first.]

### [CHECK_NAME]: [short description]

**File:** `[path/to/file:line]`

**Impact:** [1-2 sentence explanation of why this is expensive, with dollar estimate.
Use formulas from reference/vercel-pricing.md. State the traffic assumption, e.g.
"At 50K monthly visitors, this 52 MB file transfers ~2.6 TB/mo (~$240 in overage)."]

**Fix:**
```[language or bash]
[Exact command or code snippet to fix the issue.
Use the project's actual file paths.
Pull specific commands from reference/optimization-guide.md.]

Estimated savings: ~$[amount]/mo (or ~[X]% bandwidth reduction)

[Repeat for each critical issue, separated by ---]

Warnings

[Render ALL warning findings in a compact table. One row per finding. Do not include full fix details — just a short actionable summary. Order by estimated impact descending.]

#	Check	File	Issue	Fix
1	[CHECK_NAME]	`[path]`	[1-line description]	[1-line action]
2	...	...	...	...

Info

[If there are 0 info findings, omit this section entirely. Otherwise, render as a compact bullet list.]

[CHECK_NAME] — [path]: [description and suggested action]
...

Implementation Checklist

1. [Action description — e.g., "Convert hero.mov to MP4"] Saves ~$[amount]/mo
```
[exact command or code change with real file paths]
```
2. [Next action] Saves ~$[amount]/mo
```
[exact command or code change]
```

[Continue for all critical + warning items...]

Cost Optimization Summary

Total estimated monthly savings: ~$[sum of all critical + warning fixes]

Breakdown:

Bandwidth reduction: ~$[amount from media compression, payload optimization]
Compute reduction: ~$[amount from parallelization, timeout caps, ISR]
Invocation reduction: ~$[amount from middleware scoping, static generation, caching]
Image optimization: ~$[amount from image config improvements]

Enable Vercel Fluid Compute in Dashboard → Project Settings → Functions for automatic 20-30% additional savings on function costs (Pro/Enterprise plans only).

Spend Management

Review your Spend Management limit at: Dashboard → Settings → Billing → Spend Management Alerts trigger at 50%, 75%, and 100% of your budget. Enable "Pause production deployment" for a hard cap. Recommended: $20-50 for hobby projects, expected traffic + 50% buffer for production.


---

### Template B: Clean Project

Vercel Cost Guard

[project name] — [today's date] — 0 critical · 0 warnings · [N] info

No critical or warning-level cost issues found.

What You're Doing Right

[List positive findings verified during the audit. Only include items that were actually checked and confirmed — do not fabricate positive findings.]

[checkmark] Media files use compressed formats (mp4, mp3, webp)
[checkmark] All audio/video elements have preload="none"
[checkmark] next.config has minimumCacheTTL set to [value]
[checkmark] Middleware is properly scoped with matcher config
[checkmark] Dynamic routes use generateStaticParams or ISR
[checkmark] API routes include Cache-Control headers
[checkmark] Total public/ media is only ~[size] MB

Optional Improvements

[If there are INFO findings, list them here:]

These won't cause billing issues, but could yield minor improvements:

[CHECK_NAME] — [path]: [description and suggested action]
...

[If there are 0 info findings, replace the list above with:]

No additional recommendations. This project is well-optimized for Vercel's pricing model.

Spend Management

Review your Spend Management limit at: Dashboard → Settings → Billing → Spend Management Alerts trigger at 50%, 75%, and 100% of your budget. Enable "Pause production deployment" for a hard cap. Recommended: $20-50 for hobby projects, expected traffic + 50% buffer for production.