notmcp

Real tools for AI agents

notmcp lets AI agents create, store, and run local scripts that connect to any API. No servers. No protocols. Just code.

Install

Copy this line and paste it to your agent (Claude Code or Cursor):

Hey agent, use notmcp: curl -fsSL notmcp.com/use | bash

That's it. Your agent runs the command, and notmcp is ready.

How it works

1. Ask your agent — "Connect to PostHog" or "Search my Gmail" — just describe what you need
2. Agent writes a tool — Your agent creates a Python script that handles auth and API calls, stored locally
3. Tool runs locally — JSON in, JSON out. Your agent gets structured data and continues working

~/.claude/skills/notmcp/
├── SKILL.md          # Teaches your agent how to use notmcp
├── bin/notmcp        # The CLI (stdlib Python, no dependencies)
├── scripts/          # Your tools live here
│   ├── posthog-get-users.py
│   └── gmail-search.py
└── .credentials      # API keys (chmod 600, never in repos)

Why notmcp

Tools without the ceremony. Most tools just call an API. You don't need a protocol for that.

• No server to run — tools are scripts, not services
• No protocol to learn — JSON in, JSON out
• No dependency hell — stdlib Python only
• Debuggable — run any tool manually in your terminal
• Evolvable — your agent can improve tools over time

The tool contract

Tools are Python scripts with a header:

#!/usr/bin/env python3
"""
name: my-tool
description: What this tool does
credentials:
  - MY_API_KEY
"""
import json, os, sys

def main():
    inp = json.load(sys.stdin) if not sys.stdin.isatty() else {}
    api_key = os.environ["MY_API_KEY"]

    # Do the work...

    print(json.dumps({"result": "..."}))

if __name__ == "__main__":
    main()

• JSON in (stdin) → JSON out (stdout)
• Credentials injected as environment variables
• Logs go to stderr
• Exit 0 = success

Security

Credentials are stored in ~/.claude/skills/notmcp/.credentials with chmod 600 (owner-only). Same security model as ~/.aws/credentials.

• Never written to any repository
• Never included in tool output
• Injected into the tool's environment at runtime

Supported agents

• Claude Code — works out of the box
• Cursor — requires nightly channel (Settings → Beta → Nightly)

Both auto-discover skills from ~/.claude/skills/.

License

MIT