LeanProxy-MCP

The Local CLI Proxy That Slashes Your AI Token Bill

Latest Benchmark (v0.9.0)

Measured on a 3-server production-shaped MCP setup (GitHub + Garmin + Intervals.icu) with the canonical pkg/reporter.Estimator (1 token ≈ 4 chars). Reproduce with make bench.

All numbers above come from tests/bench/token_economy_bench_test.go and pkg/reporter/cost.go. Full results and methodology: docs/benchmark-results.md.

The MCP Schema Tax is Killing Your AI Budget

Every MCP server you connect injects thousands of tokens into every LLM request — even when you never use it. This is the "Schema Tax":

flowchart LR
    IDE["Your IDE"] --> MCP["MCP Gateway"]

    subgraph MCP["MCP Gateway"]
        S1["GitHub (41 tools)"]
        S2["Garmin (100 tools)"]
        S3["Intervals.icu (10 tools)"]
    end

    MCP --> LLM["LLM Provider"]

    note1["~4,570 tokens"]
    note2["~11,130 tokens"]
    note3["~1,130 tokens"]
    total["TOTAL: ~16,830 tokens"]

    S1 -.-> note1
    S2 -.-> note2
    S3 -.-> note3
    note1 --- total
    note2 --- total
    note3 --- total
    total -.-> LLM

    style MCP fill:#ff6b6b,color:#fff
    style LLM fill:#ee5a5a,color:#fff

The result? You're burning tokens on tool definitions you'll never use in that session. Numbers above are measured on v0.9.0 — see Latest Benchmark.

Enter LeanProxy: Your Token Firewall

LeanProxy sits between your IDE and MCP servers as a smart gateway. It loads tool schemas only when needed — reducing the schema tax to a single 158-token router payload.

flowchart LR
    IDE["Your IDE"] --> Gateway["LeanProxy Gateway"]

    subgraph Gateway["LeanProxy Gateway"]
        Router["Router: 3 tools (~158 tokens)"]
        JIT["JIT Schema Loading"]
        Cache["Automatic Caching"]
        Firewall["Token Firewall"]
        Pool["Connection Pooling"]
    end

    Gateway -.->|"loads on demand"| GH["GitHub"]
    Gateway -.->|"loads on demand"| Garmin["Garmin"]
    Gateway -.->|"loads on demand"| Intervals["Intervals.icu"]

    Router --> Firewall --> Pool --> GH
    Router --> Firewall --> Pool --> Garmin
    Router --> Firewall --> Pool --> Intervals

    Router --> JIT
    JIT --> Cache

    style Gateway fill:#1a1a2e,color:#fff
    style Router fill:#00ADD8,color:#fff
    style JIT fill:#00ADD8,color:#fff
    style Cache fill:#00ADD8,color:#fff
    style Firewall fill:#00ADD8,color:#fff
    style Pool fill:#00ADD8,color:#fff

Real Results, Real Savings

86-99% Token Reduction in Production Sessions (Measured v0.9.0)

The Math Doesn't Lie

Measured on v0.9.0 with the same MCP server tool counts as production. Native MCP + 100% cache hit still costs you at 0.25x (cache read isn't free!). We use the raw tools/list token count for Native and the same pkg/reporter.Estimator (1 token ≈ 4 chars) for both columns.

The earlier "4 servers" row is no longer applicable — the Stitch MCP server is no longer available, so the canonical production shape is 3 servers. See docs/benchmark-results.md for the full table and methodology.

Key Features

Quick Start

One-Line Install

# macOS/Linux via Homebrew
brew tap mmornati/leanproxy-mcp https://github.com/mmornati/leanproxy-mcp
brew install leanproxy-mcp

# ...or download binary for your platform
curl -fsSL https://github.com/mmornati/leanproxy-mcp/releases/latest/download/leanproxy-mcp.tar.gz | tar xz

Configure Your IDE

Add LeanProxy as an MCP server in your opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "leanproxy": {
      "type": "local",
      "command": ["leanproxy-mcp", "server", "run", "--stdio"],
      "enabled": true
    }
  }
}

Run It

# Start with your MCP servers
leanproxy-mcp server run --stdio

# Preview savings without executing
leanproxy-mcp server run --dry-run --stdio

# Generate a detailed savings report
leanproxy-mcp report --output report.md

Architecture

flowchart TB
    subgraph IDE["Your IDE"]
        Client[(MCP Client)]
    end

    subgraph Gateway["LeanProxy Gateway"]
        Router["Router<br/>(3 tools, ~158 tokens)"]
        JIT["JIT Schema Cache"]
        Firewall["Token Firewall<br/>(Secret Redaction)"]
        Pool["Connection Pool<br/>& Circuit Breaker"]
        Federation["Federation<br/>(Multi-org)"]
    end

    subgraph Servers["MCP Servers"]
        GH["GitHub<br/>(stdio)"]
        Garmin["Garmin<br/>(HTTP)"]
        Intervals["Intervals.icu<br/>(SSE)"]
    end

    Client <--> Router
    Router <--> JIT
    Router <--> Firewall
    Firewall <--> Pool
    Pool <--> Federation

    Pool --- GH
    Pool --- Garmin
    Pool --- Intervals

    JIT -.->|"loads on call"| GH
    JIT -.->|"loads on call"| Garmin
    JIT -.->|"loads on call"| Intervals

    style Gateway fill:#1a1a2e,color:#fff
    style Router fill:#00ADD8,color:#fff
    style JIT fill:#00ADD8,color:#fff
    style Firewall fill:#00ADD8,color:#fff
    style Pool fill:#00ADD8,color:#fff
    style Federation fill:#00ADD8,color:#fff

v0.8.0: What's New

Join the Community

GitHub • Documentation • Issues

License

MIT © Marco Mornati