rtk-mcp

MCP server bridge for RTK (Rust Token Killer) — token-optimized CLI output for any MCP-compatible client.

What it does

rtk-mcp exposes a single MCP tool (run_command) that routes shell commands through RTK for 60-90% token reduction before the output reaches your LLM's context window.

MCP Client (Claude Desktop, Cursor, Windsurf, ...)
  → rtk-mcp (this server)
    → rtk git status     ← filtered output (78% fewer tokens)
    → returns to LLM

Without RTK installed, commands execute normally (no filtering, no token savings).

Why

RTK already saves tokens for Claude Code and Gemini CLI via hooks. But hooks are client-specific — each new AI tool needs its own integration.

MCP is a universal protocol. One server, every client:

Client	MCP Support
Claude Desktop	Yes
Cursor	Yes
Windsurf	Yes
Cline (VS Code)	Yes
Continue	Yes
Zed	Yes
VS Code (native)	Yes
GitHub Copilot	Yes

Real-world savings

Measured over 25 days of daily usage with RTK:

Filter	Token reduction
cargo test	97.8%
env	99.3%
cargo clippy	92.5%
find	79.2%
ls	67-78%
grep	64.4%

Total: 5.3M tokens saved across 4,876 commands.

Install

Prerequisites

Install RTK first:

cargo install --git https://github.com/rtk-ai/rtk

# Verify
rtk --version   # Should show "rtk X.Y.Z"
rtk gain         # Should work (not "command not found")

Build rtk-mcp

git clone https://github.com/ousamabenyounes/rtk-mcp.git
cd rtk-mcp
cargo build --release

The binary is at target/release/rtk-mcp.

Configure your MCP client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "rtk": {
      "command": "/path/to/rtk-mcp"
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "rtk": {
      "command": "/path/to/rtk-mcp"
    }
  }
}

VS Code / Windsurf / Cline

Add to your MCP settings (check each client's documentation for the exact config file path):

{
  "mcpServers": {
    "rtk": {
      "command": "/path/to/rtk-mcp"
    }
  }
}

Usage

Once configured, the run_command tool is available in your AI assistant. It accepts:

Parameter	Type	Required	Description
command	string	Yes	The command to execute (e.g. git status, cargo test)
cwd	string	No	Working directory (defaults to server cwd)

Example tool calls from the LLM:

{"name": "run_command", "arguments": {"command": "git log --oneline -5", "cwd": "/my/project"}}
{"name": "run_command", "arguments": {"command": "cargo test"}}
{"name": "run_command", "arguments": {"command": "ls -la src/"}}

Security

Command allowlist

Only allowlisted commands are accepted. Dangerous commands like bash, sh, rm, sudo are blocked:

Allowed: git, cargo, npm, npx, pnpm, pytest, ruff, mypy, pip, uv, go, golangci-lint, docker, grep, find, ls, cat, head, tail, wc, env, echo, pwd, gh, curl, wget, node, tsc, next, prettier, eslint, biome, playwright, prisma, vitest, dotnet, psql, make, tree

Blocked: Everything else (bash, sh, rm, sudo, chmod, python, etc.)

Other protections

• Shell parsing: Uses shlex for proper quoted argument handling (no split_whitespace injection)
• Length limit: Commands capped at 4096 characters
• No shell invocation: Uses Command::new() directly, never spawns a shell
• RTK validation: Verifies the correct RTK binary is installed at startup (avoids name collision)
• Exit code propagation: Failed commands return isError: true in MCP response

How it works

┌──────────────┐     stdio (JSON-RPC)     ┌──────────────┐
│  MCP Client  │ ◄──────────────────────► │   rtk-mcp    │
│  (Cursor,    │                          │              │
│   Claude     │                          │  1. Parse    │
│   Desktop)   │                          │  2. Validate │
│              │                          │  3. Execute  │
└──────────────┘                          └──────┬───────┘
                                                 │
                                          ┌──────▼───────┐
                                          │     rtk      │
                                          │  (filtering) │
                                          │              │
                                          │ git status   │
                                          │ → 78% less   │
                                          │   tokens     │
                                          └──────────────┘

1. MCP client sends a tools/call request with a command string
2. rtk-mcp validates the command against the allowlist
3. Parses arguments with shlex (handles quotes correctly)
4. Executes via rtk <command> for filtered output
5. If RTK is unavailable, falls back to raw command execution
6. Returns output with exit code information

Development

# Run tests
cargo test

# Build
cargo build --release

# Test MCP protocol manually
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.1"}}}' | ./target/release/rtk-mcp

Credits

• RTK (Rust Token Killer) by rtk-ai — all filtering logic
• rmcp — Rust MCP SDK
• Built with Claude Code

License

MIT