Your entire codebase in Claude's context. At 1/5th the cost.

mcp-injector is a local MCP daemon that compresses your codebase before sending it to Claude Code, Cursor, or VS Code — using AST body folding and canonical determinism for maximum prompt cache hits.

Install Free → Run Benchmark

Free for codebases under 100,000 lines. No signup required.

Real numbers on real codebases

Run mcp-benchmark on your own repo to see your exact savings.

🐍

Django Source

Total files scanned 2,359
Raw context tokens 5,554,622
Compressed context 1,524,435
Reduction 72.6%
Saved per run $12.09
At 10 runs/day $3,627/mo

Next.js Source

Total files scanned 21,848
Raw context tokens 23,852,376
Compressed context 4,967,583
Reduction 79.2%
Saved per run $56.65
At 10 runs/day $16,996/mo

* Based on Claude Sonnet pricing: $3.00/1M input tokens. Compression tier 3 (fold + strip). Reproducible: clone any repo above and run mcp-benchmark yourself.

How it works

📁

Pre-indexes your codebase

Runs as a persistent local daemon. Builds a SQLite index via file watchers — inotify on Linux, FSEvents on Mac. First query is sub-millisecond on any size codebase.

🗜️

AST body folding

Strips function bodies, keeps signatures and imports. Supports Go, Python, TypeScript, JavaScript, Java, Rust. Compression is fully reversible — Claude can retrieve full source when needed.

🎯

Canonical determinism

Files sorted alphabetically, volatile metadata stripped. Output is byte-identical on every run so Anthropic's KV prompt cache fires instead of invalidating. This is the feature other compressors miss.

💰

You save money

Fewer tokens sent. More cache hits. Two separate cost levers working simultaneously. No cloud, no telemetry, runs entirely on your machine.

Get started in 30 seconds

Mac / Linux
Windows
curl -fsSL https://foldwork.dev/install | sh

Automatically detects and configures: Claude Desktop, Cursor, VS Code

🤖 Claude Desktop
⚡ Cursor
💻 VS Code (Continue)
{ "mcpServers": { "mcp-injector": { "command": "/usr/local/bin/mcp-injector", "env": { "MCP_WORKSPACE": "${workspaceFolder}" } } } }

Simple pricing

Free for personal projects. Pay only when it makes financial sense.

Monthly Yearly (save 31%)

Free

$0 / forever
  • Codebases under 100,000 lines
  • All languages (Go, Python, TS, JS, Java, Rust)
  • get_project_map tool
  • Basic CCR caching
  • No signup, no license key
Download Free →

Common questions

No. mcp-injector runs entirely on your machine. Your code never leaves your computer. The daemon reads files locally, compresses them locally, and serves them to your IDE via stdio transport. No cloud, no telemetry, no external API calls.

Headroom compresses context on demand. mcp-injector pre-indexes your codebase into a SQLite catalog so the first query is sub-millisecond on any size repo. More importantly, mcp-injector guarantees byte-identical prompt output on every run through canonical determinism — alphabetical file sorting and metadata stripping — so Anthropic's KV prompt cache fires on every session. Headroom's on-demand architecture makes this impossible. You save tokens AND maximise cache hits simultaneously.

Claude Desktop, Cursor, and VS Code with Continue. The installer auto-detects which are installed and configures them automatically.

Go, Python, TypeScript, JavaScript, Java, and Rust. Other file types pass through unmodified.

mcp-injector exposes an injector_retrieve tool. When Claude needs the complete implementation of a compressed function, it calls this tool to fetch the full original source. Compression is always reversible.

Yes. Pro includes 3 seats. Each developer runs their own local daemon — there's no shared server to manage.