Open Source MCP Server

Your AI reads too much.
Fix that.

C3 is a local intelligence layer that gives AI coding tools surgical project understanding instead of brute-force file reads.

Get Started See the Numbers

your AI — with C3

AI> c3_search(query="auth middleware") 3 files · 0.04s · 12 tokens AI> c3_compress("services/auth.py", mode="map") 340 lines → 28 token structural map AI> c3_read("services/auth.py", symbols=["verify_token"]) 22 lines extracted · exactly what was needed # Without C3: 3 full files, ~4,200 tokens # With C3: 3 targeted calls, ~180 tokens

Setup

Three commands. Zero config.

Python 3.10+ is the only requirement. No cloud, no API keys. Runs 100% locally.

Clone & Install

Grab the repo and install dependencies.

git clone && pip install -r requirements.txt

Initialize

C3 indexes your code, detects your IDE, and registers the MCP server.

c3 init /path/to/project

Code

Open your AI tool. C3 tools appear automatically via MCP protocol.

claude / codex / gemini / cursor

Benchmarks

Measured, not marketed

14 tasks across 7 categories. Same model, same codebase. Only C3 changes the outcome.

Code Review

+50pp

100% win rate · C3 0.71 vs Base 0.20

Bug Detection

+36pp

Structural maps reveal hidden patterns

Refactoring

+31pp

100% win rate · baseline timed out

Call Chains

0.90

Near-perfect cross-file tracing

Session Net Savings

76%

After C3 overhead (mandates, schemas, recalls)

Avg Score

0.74

C3 0.736 vs Baseline 0.575

Model: Claude Sonnet 4.6 · Tasks: 14 · C3 Wins: 8/14 · C3 21% faster Full results on GitHub →

Toolkit

Six tools your AI didn't know it needed

c3_search

TF-IDF + optional semantic search via Ollama. Finds files by meaning, not just text. Sub-second results.

c3_compress

Structural maps that expose every class, function, and signature. 340 lines becomes a 28-token overview.

c3_read

Symbol-level extraction. Pull one function from a 2,000-line file. Read 22 lines instead of 2,000.

c3_validate

Native syntax checkers for 15+ languages. Background validation with result caching.

c3_memory

Persistent fact store with auto-learning. Your AI remembers decisions and conventions across sessions.

c3_filter

Pre-processes logs, terminal output, and data files. Extracts signal from noise before it hits context.

Compatibility

Works everywhere MCP does

One install covers every AI coding tool that supports Model Context Protocol.

Tool	Provider	MCP Tools	Hooks	Instruction Sync	Notes
C Claude Code	Anthropic	✓	✓	✓	Full support — hooks, CLAUDE.md sync, budget management
VS VS Code	Microsoft	✓	—	—	MCP tools via Copilot extension
Cx Codex CLI	OpenAI	✓	—	✓	MCP tools + AGENTS.md + session config
G Gemini CLI	Google	✓	—	✓	MCP tools + GEMINI.md + session config

Support C3

Free and open source. If C3 saves you tokens, extends your sessions, and makes AI coding better — consider sponsoring its development.

Star on GitHub

$ git clone https://github.com/drknowhow/C3.git $ pip install -r requirements.txt $ c3 init /path/to/your/project Indexed 847 files · MCP registered · ready

Your AI reads too much.Fix that.