notmcp

You have access to a local toolbox of executable scripts. These tools let you interact with external APIs and services on behalf of the user.

Discovering Tools

To see what tools are available:

~/.claude/skills/notmcp/bin/notmcp list

To search for tools by keyword:

~/.claude/skills/notmcp/bin/notmcp search <query>

Running Tools

Run tools with JSON input:

~/.claude/skills/notmcp/bin/notmcp run <tool-name> --input '{"key": "value"}'

Tools return JSON to stdout. Exit code 0 means success.

If a tool doesn't need input, you can omit the --input flag:

~/.claude/skills/notmcp/bin/notmcp run <tool-name>

Handling Credentials

If a tool needs credentials (API keys, tokens), it will fail with an error like:

Error: Missing credential(s): POSTHOG_API_KEY

When this happens:

1. •Ask the user for the credential value
2. •Store it securely:

echo "the-secret-value" | ~/.claude/skills/notmcp/bin/notmcp creds set CREDENTIAL_NAME

To see what credentials are already stored:

~/.claude/skills/notmcp/bin/notmcp creds list

Connecting to Services

When the user wants to connect to a service, guide them interactively, one step at a time. Wait for the user to complete each step before proceeding to the next.

Connection Process (Interactive)

Step 1: Open the credentials page

• •Run open "URL" to open the page in their browser
• •Tell them what page opened and wait for confirmation they see it

Step 2: Guide them through the UI (one action at a time)

• •Give ONE instruction, then wait for them to do it
• •Don't dump all steps at once - be conversational
• •Example: "Click 'Generate new token'" → wait → "Now copy the token that appears"

Step 3: Collect the credential

• •Ask them to paste the token/key
• •Store it immediately when they provide it

Step 4: Verify silently

• •Run a verification API call
• •Only tell them if it fails - if it works, just confirm "Connected!"

Important: Be Interactive

BAD (dumping everything):

Here's how to connect:
1. Go to URL
2. Click X
3. Click Y  
4. Copy Z
5. Paste it here

GOOD (interactive):

I'll open the GitHub tokens page for you.
[opens browser]
Let me know when you see the page.

[user: "ok I see it"]

Click "Generate new token" at the top.

[user: "done"]

Now copy the token that appears and paste it here.

Verifying Connections

After storing credentials, verify they work by making a simple API call:

• •Most REST APIs have a /me, /user, or /auth endpoint
• •Use the http-get tool or make a quick urllib request
• •If it fails, troubleshoot with the user before proceeding
• •If it works, just say "Connected!" - don't over-explain

Common Services

Google (Gmail, Drive, Calendar) - App Password

• •URL: https://myaccount.google.com/apppasswords
• •Requires 2FA enabled on the account
• •Guide: "Select 'Other (Custom name)', enter 'notmcp', click Generate"
• •Store as: GOOGLE_APP_PASSWORD and GOOGLE_EMAIL

GitHub - Personal Access Token

• •URL: https://github.com/settings/tokens/new?description=notmcp&scopes=repo,user
• •Guide: "Click 'Generate token' and copy it"
• •Store as: GITHUB_TOKEN
• •Verify: GET https://api.github.com/user with Authorization header

Slack - Bot Token

• •URL: https://api.slack.com/apps
• •Guide: "Create New App → From scratch → OAuth & Permissions → Install to Workspace"
• •Store as: SLACK_BOT_TOKEN
• •Verify: POST https://slack.com/api/auth.test

Notion - Integration Token

• •URL: https://www.notion.so/my-integrations
• •Guide: "New integration → Name 'notmcp' → Copy the Internal Integration Secret"
• •Store as: NOTION_TOKEN
• •Remind user to share specific pages with the integration

OpenAI - API Key

• •URL: https://platform.openai.com/api-keys
• •Guide: "Create new secret key → Copy it"
• •Store as: OPENAI_API_KEY
• •Verify: GET https://api.openai.com/v1/models

Linear - API Key

• •URL: https://linear.app/settings/api
• •Store as: LINEAR_API_KEY

Stripe - Secret Key

• •URL: https://dashboard.stripe.com/apikeys
• •Store as: STRIPE_SECRET_KEY

Other Services

For services not listed:

1. •Search for their API or Developer documentation
2. •Find where to create API keys or tokens
3. •Guide the user through the process
4. •Store with a descriptive name: {SERVICE}_API_KEY
5. •Verify with a simple API call before building tools

Using Context7 for API Documentation (Optional)

Context7 provides up-to-date API documentation. If CONTEXT7_API_KEY is set, fetch current docs before creating tools to avoid using outdated or hallucinated endpoints:

~/.claude/skills/notmcp/bin/notmcp run context7-docs --input '{"library": "googleapis/gmail", "topic": "send"}'

To set up Context7:

1. •Open: https://context7.com/dashboard
2. •Sign up (free) and copy your API key
3. •Store: echo "xxx" | ~/.claude/skills/notmcp/bin/notmcp creds set CONTEXT7_API_KEY

Without Context7, you can still create tools using your knowledge, but results may be less accurate for newer APIs.

Creating Tools

Only create new tools when the user explicitly asks (e.g., "save this as a tool", "make this reusable", "create a tool for this").

To create a new tool:

~/.claude/skills/notmcp/bin/notmcp create tool-name

This creates a template at ~/.claude/skills/notmcp/scripts/tool-name.py.

Then edit the script to implement the tool logic. Follow these conventions:

Tool Contract

• •Input: JSON via stdin (parsed with json.load(sys.stdin))
• •Output: JSON to stdout (use print(json.dumps(result)))
• •Logs: Write debug info to stderr
• •Exit code: 0 for success, nonzero for failure
• •Dependencies: Use Python stdlib only (urllib.request, json, os, etc.)

Tool Header Format

Every tool must have a docstring header declaring its metadata:

#!/usr/bin/env python3
"""
name: tool-name
description: What this tool does (one line)
credentials:
  - API_KEY_NAME
  - ANOTHER_SECRET
input:
  param1: string (required)
  param2: int (optional, default 10)
output:
  result: description of output
"""

Example Tool Structure

#!/usr/bin/env python3
"""
name: example-api
description: Fetch data from Example API
credentials:
  - EXAMPLE_API_KEY
input:
  query: string (required)
output:
  results: list of matching items
"""

import json
import os
import sys
from urllib.request import urlopen, Request

def main():
    # Read input
    inp = json.load(sys.stdin) if not sys.stdin.isatty() else {}
    
    # Get credentials (injected by notmcp run)
    api_key = os.environ["EXAMPLE_API_KEY"]
    
    # Make API call
    query = inp.get("query", "")
    req = Request(f"https://api.example.com/search?q={query}")
    req.add_header("Authorization", f"Bearer {api_key}")
    
    response = urlopen(req)
    data = json.loads(response.read())
    
    # Return result
    print(json.dumps({"results": data["items"]}))

if __name__ == "__main__":
    main()

Best Practices

1. •Handle pagination - Don't return unbounded results. Implement limits and cursors.
2. •Handle rate limits - Add delays or retry logic for APIs with rate limits.
3. •Return compact output - Summarize large responses. Agents work better with concise data.
4. •Fail gracefully - Return {"error": "message"} with a helpful error description.
5. •Use stdlib - Avoid pip dependencies so tools are portable and self-contained.