Skip to content

compatibility.md

Latest commit

 

History

History
165 lines (132 loc) · 4.94 KB

compatibility.md

File metadata and controls

165 lines (132 loc) · 4.94 KB
Raw

Works with the agents you already use

pin-the-protocol adds nothing new to the wire. The gateway is a standard MCP server over stdio, and skill-lock is a standard Agent Skills content check. So any agent that already speaks MCP or loads Skills can adopt it without code changes:

  • To protect an MCP server, point the agent at the gateway instead of the server. The gateway takes the real server command after --.
  • To protect a skill, run skill-lock verify <dir> before the skill activates (a pre-activation hook, a wrapper, or a CI gate).

Approve once, from this repo, before wiring an agent up:

./demo/demo.sh                       # or, minimally:
demo/scripts/trust-store.sh start
node demo/client/approve.mjs weather.v1

MCP: wrap any server with the gateway

The pattern is always the same — replace <server cmd> with the command the agent runs today, and give the pin a name:

node /abs/path/pin-the-protocol/demo/gateway/gateway.mjs --pin <name> -- <server cmd>

Claude Code / Claude Desktop — mcpServers

docs

{
  "mcpServers": {
    "weather": {
      "command": "node",
      "args": ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
               "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"],
      "env": { "TRUST_STORE_URL": "http://127.0.0.1:8080" }
    }
  }
}

goose — ~/.config/goose/config.yaml

docs

extensions:
  weather:
    type: stdio
    cmd: node
    args: ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
           "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"]

Cursor — .cursor/mcp.json

docs

{
  "mcpServers": {
    "weather": {
      "command": "node",
      "args": ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
               "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"]
    }
  }
}

OpenClaw — ~/.openclaw/config

docs

{
  "mcp": {
    "servers": {
      "weather": {
        "command": "node",
        "args": ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
                 "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"]
      }
    }
  }
}

Hermes Agent (Nous Research)

docs

{
  "mcpServers": {
    "weather": {
      "command": "node",
      "args": ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
               "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"]
    }
  }
}

VS Code / GitHub Copilot — .vscode/mcp.json

docs

{
  "servers": {
    "weather": {
      "type": "stdio",
      "command": "node",
      "args": ["/abs/path/pin-the-protocol/demo/gateway/gateway.mjs", "--pin", "weather.v1",
               "--", "node", "/abs/path/pin-the-protocol/vendor/weather-server.mjs"]
    }
  }
}

Gemini CLI, OpenCode, Amp, Roo Code, Kiro, and other MCP clients

Every MCP client exposes a "command + args" server entry. Use the same shape: command: node, args: [ …/gateway.mjs, --pin, <name>, --, <server cmd> ]. The gateway is transparent — the agent sees exactly the upstream tools, minus any that failed verification.

Skills: verify before activation, on any Skills-capable agent

Agent Skills are supported across Claude, goose, Cursor, Copilot, Gemini CLI, OpenHands, Letta, and many more. skill-lock is agent-agnostic because it checks the SKILL.md bytes, not the agent:

# one-time, per skill:
node demo/skill-lock/skill-lock.mjs approve ~/.claude/skills/release-notes

# before activation (hook, wrapper, or CI gate):
node demo/skill-lock/skill-lock.mjs verify ~/.claude/skills/release-notes || {
  echo "skill changed since approval — refusing to activate"; exit 1;
}

Exit code 0 means verified; 1 means drift (with a diff printed); 3 means not yet approved. Wire it into whatever runs before your agent loads a skill — a shell wrapper, a pre-commit hook for a shared skills repo, or a CI check that gates skill updates.

Verify it yourself with MCP Inspector

Point the official MCP Inspector at the vendor server to see the original vs. altered tool descriptions directly:

npx @modelcontextprotocol/inspector node vendor/weather-server.mjs           # original
POISON=1 npx @modelcontextprotocol/inspector node vendor/weather-server.mjs   # altered

Then point it at the gateway and watch the altered variant get blocked.