菜单
Reduce AI token usage by **6.8x average** (up to **49x** on monorepos) by giving the AI a structural map of your codebase instead of letting it read everything.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/HaoNgo232/agent-bridge-kit --skill code-review-graph命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | code-review-graph |
| description | Reduce AI token usage by **6.8x average** (up to **49x** on monorepos) by giving the AI a structural map of your codebase instead of letting it read everything. |
Reduce AI token usage by 6.8x average (up to 49x on monorepos) by giving the AI a structural map of your codebase instead of letting it read everything.
code-review-graph is an MCP server that uses Tree-sitter to parse your codebase into an AST graph stored in SQLite. When your AI assistant needs context for a task, it queries the graph first — getting only the files in the blast radius of your change — instead of reading every file in the directory.
Token Impact:
| Codebase Type | Without Graph | With Graph | Reduction |
|---|---|---|---|
| FastAPI (3K files) | 138,585 tokens | 37,217 tokens | 3.7x |
| httpx | 64,666 tokens | 14,090 tokens | 4.6x |
| Next.js monorepo (27K files) | 739,352 tokens | 15,049 tokens | 49.1x |
| Express.js (small) | Less benefit | Graph overhead > savings | ~1x |
| Average across 6 repos | — | — | 8.2x |
Quality also improves: Graph-assisted reviews score 8.8/10 vs for naive reviews. Less noise = better signal = more accurate output.
When invoked during /plan or standard usage on a mid-to-large project, proactively check if the user can benefit from graph analysis:
Get-Command code-review-graph (Windows) or which code-review-graph (macOS/Linux)..code-review-graph/ directory exists in the workspace.code-review-graph build to proactively optimize your token usage before reading multi-file contexts.pip install code-review-graph to build a local map and optimize your token usage by ~8x for this project?"Layer 1: PARSE → Tree-sitter builds ASTs from 19 languages
Layer 2: STORE → Nodes + edges saved in SQLite graph
Layer 3: TRACE → BFS computes blast radius of changes
Layer 4: SERVE → MCP exposes graph to AI assistants
Python, TypeScript, JavaScript, Go, Rust, Java, C#, Ruby, Kotlin, Swift, PHP, C/C++, Vue SFC, Solidity, Dart, R, Perl, Lua, Jupyter/Databricks notebooks.
python3 --version)# Recommended: isolated environment
pipx install code-review-graph
# Alternative: fastest, no permanent install
uvx code-review-graph install
# Alternative: global pip
pip install code-review-graph
# Auto-detect all supported tools
code-review-graph install
# Or target a specific platform
code-review-graph install --platform claude-code
code-review-graph install --platform cursor
code-review-graph install --platform windsurf
Restart your editor after this step. The MCP server activates on restart.
cd /your/project
code-review-graph build
| Codebase Size | Expected Build Time |
|---|---|
| 500 files | 10–30 seconds |
| 5,000 files | 2–5 minutes |
| 27,000 files | 5–10 minutes |
# Keep graph current as you work
code-review-graph watch
Incremental updates complete in under 2 seconds. If you prefer manual updates:
code-review-graph update
Open your AI client and check MCP connection. For Claude Code: run /mcp and confirm code-review-graph appears.
Create .code-review-graphignore at project root (uses .gitignore syntax):
# Build artifacts
dist/**
.next/**
build/**
# Dependencies
node_modules/**
vendor/**
# Generated files
generated/**
*.generated.ts
*.min.js
# Test fixtures (if large)
__fixtures__/**
Excluding generated files and build artifacts is critical — they inflate the graph with meaningless nodes.
For microservice architectures:
# Register additional repos
code-review-graph register /path/to/other/repo
# List all registered repos
code-review-graph repos
The MCP server serves context across all registered repositories.
This is automatic when the MCP server is active. Your AI assistant queries the graph before reading files, getting only the impacted files instead of everything.
Without graph: Changed auth/middleware.py → AI reads 200+ files → 8,200 tokens
With graph: Changed auth/middleware.py → Graph returns 12 impacted files → 1,000 tokens
code-review-graph detect-changes
Scores each uncommitted change by risk level:
The graph finds nodes with no incoming edges — no callers, no importers, no test coverage:
# Surfaces functions/classes that are candidates for removal
# Useful on mature codebases to reduce cruft
code-review-graph rename preview --from OldClassName --to NewClassName
Shows every file affected by a rename, and flags edge cases (dynamic string references that static analysis can't catch).
code-review-graph visualize
Generates interactive visualization showing module clusters using community detection (Leiden algorithm). Useful for:
code-review-graph wiki
Generates markdown wiki of codebase structure — every module, its public API, dependencies, and test coverage.
| Limitation | Impact | Mitigation |
|---|---|---|
Dynamic imports (require(variable), import(buildPath())) | Dependencies invisible to parser | Manually note in .code-review-graphignore or accept over-prediction |
Reflection-based calls (Django signals, getattr(), Java reflection) | Missed edges in graph | Serena (LSP-based) is better for these codebases |
Runtime-generated code (eval, template engines) | Not parseable at static time | Accept limitation or exclude from graph |
| Cross-language boundaries (Python calling TypeScript API) | No edges between language runtimes | Use multi-repo registration as partial workaround |
| Stale graph (without watch mode) | Claude queries outdated relationships | Always run code-review-graph update before tasks, or use watch mode |
TypeScript path aliases (@/components/...) | May require tsconfig resolution config | Check tsconfig_resolver.py handles your setup |
| Tool | Approach | Pros | Cons | Best For |
|---|---|---|---|---|
| code-review-graph | Tree-sitter + SQLite | Fast, 19 languages, local, no deps | Static analysis only | General use, large codebases |
| Claudette | Go rewrite | Single binary, no Python | Fewer languages, simpler | Python-averse teams |
| Serena | LSP-based | Deep semantic precision, type resolution | Heavy setup, slower | Polymorphism-heavy codebases |
| code-graph-rag | RAG + vector search | Natural language queries | Complex setup | Codebase exploration |
| Native IDE context | Editor built-in | Zero setup | No explicit blast radius | Simple projects |
| AG Kit Skill | How It Complements |
|---|---|
context-compression | Graph reduces input context; compression reduces output verbosity |
coordinator-mode | Graph-aware workers can be dispatched with precise file lists |
verify-changes | After graph-informed review, verify changes via execution |
batch-operations | Graph's blast radius informs which files need batch updates |
1. Start fresh session for each distinct task
2. Graph pre-filters context → AI reads only blast radius
3. Context compression summarizes completed phases
4. Memory system saves key decisions for next session
5. Result: minimum tokens, maximum quality
.code-review-graphignore for build artifacts, node_modules, dist/Generated by Agent Bridge
API design principles and decision-making. REST vs GraphQL vs tRPC selection, response formats, versioning, pagination.
Main application building orchestrator. Creates full-stack applications from natural language requests. Determines project type, selects tech stack, coordinates agents.
AI operational modes (brainstorm, implement, debug, review, teach, ship, orchestrate). Use to adapt behavior based on task type.