sense

name	sense
description	Local-only MCP server that gives AI coding agents structural understanding of your codebase — symbol graph, blast radius, semantic search, and auto-detected project conventions. Reduces tool calls and tokens for any task that depends on understanding code relationships rather than reading text.
license	Other

Sense: Codebase Understanding for AI Coding Agents

Sense is an MCP server that gives AI agents the structural understanding of a codebase that a senior engineer carries in their head. It runs as one Go binary with a local SQLite index — no API keys, no SaaS, no cloud calls.

When to Use This Skill

Use Sense whenever the task depends on understanding code relationships, not just reading text:

• "Who calls this function?" / "What does this function call?"
• "What breaks if I change this signature?" (blast radius)
• "Find the implementation of the X pattern" (semantic search)
• "What conventions does this project follow?" (e.g., naming, Rails idioms, Go interface patterns)
• "Find dead code"
• "Show me all routes / handlers / models / controllers"

Do NOT use Sense for:

• Locating a string literal or log message — plain grep is faster.
• Editing code — Sense is read-only by design.

Setup (one-time per project)

# Install the binary (macOS / Linux):
curl -fsSL https://luuuc.github.io/sense/install.sh | sh

# In the project root:
sense scan      # builds the .sense/ index (tree-sitter + embeddings)
sense setup     # writes .mcp.json + CLAUDE.md routing + hooks

After sense setup, your AI tool (Claude Code, Cursor, Codex CLI) has Sense's four tools available via MCP.

Tools

Tool	Capability
sense_graph	Symbol relationships — callers, callees, inheritance, tests, dead code
sense_search	Hybrid semantic + keyword search with text fallback
sense_blast	Blast radius, affected code, affected tests, risk score
sense_conventions	Auto-detected project conventions (Rails, React, Django, Go, naming)

Plus a generated .sense/summary.md at session-start: top namespaces, hub symbols, entry points, conventions — so the AI is oriented before its first tool call.

How to Call (MCP)

After sense setup, the tools are exposed automatically. Prefer them over grep/glob for any structural question:

• "Who calls X?" → sense_graph symbol="X" direction=callers
• "What would break if I change Y?" → sense_blast symbol="Y"
• "Find code that does Z" → sense_search query="natural language description"
• "How does this project define services?" → sense_conventions

Performance

Operation	p50 / p95
Graph query	0.2 ms / 3 ms
Blast radius	0.1 ms / 10 ms
Conventions	16 ms / 16 ms
Cold start	48 ms
Full scan	4.9 s
Incremental scan	2.3 s

Measured on Sense's own codebase (382 files, 4,032 symbols).

Why Sense Helps Agents

Across 7 real-world codebases (Discourse, Flask, Next.js, Axum, Gin, Javalin, and a private e-commerce repo), Claude Code with Sense used 47% fewer tool calls and 32% fewer tokens per task at the same correctness as baseline Claude Code. Against 7 other code-intelligence MCPs (Serena, Probe, GitNexus, GrepAI, ChunkHound, codebase-memory-mcp, TokenSave, Roam), Sense ranks #1 on the head-to-head leaderboard.

Sense doesn't make the model smarter. It gives the model structural understanding so it stops wasting effort.

Language Support

Full tier (framework inference): Ruby + Rails (associations, callbacks, routes, Stimulus, Turbo), TypeScript/JavaScript + React (JSX), Python + Django (models, URLs) + FastAPI (routes, Depends), Go, Rust, ERB.

Standard tier (symbols + calls + inheritance + imports + blast + search): Java, Kotlin, C#, C++, C, PHP, Scala.

Where to Learn More

• Repo: https://github.com/luuuc/sense
• Homepage: https://luuuc.github.io/sense/
• Benchmark methodology: https://github.com/luuuc/sense/tree/main/bench
• Head-to-head leaderboard: https://github.com/luuuc/sense/blob/main/docs/bench-leaderboard.svg
• Adding a language: https://github.com/luuuc/sense/blob/main/CONTRIBUTING-A-LANGUAGE.md
• CLI reference: https://github.com/luuuc/sense/blob/main/CLI.md