AI Workbench MCP

A tiny, read-only MCP server for reusable AI prompts and assistant blueprints.

AI Workbench MCP exposes a small local catalog over Model Context Protocol stdio. It is intentionally boring in the best way: no network calls, no shell execution, no account access, no file writes, no hidden provider request.

It gives an MCP host three tools:

Why this exists

A lot of AI repos jump straight from "here is a prompt" to "this is an agent." I wanted a smaller boundary that is easy to inspect.

The server keeps the useful parts local and makes its limits obvious:

• read-only tool contracts;
• explicit MCP trust hints;
• bounded input sizes;
• strict top-level schemas;
• no runtime dependencies outside the Python standard library;
• real stdio handshake tests;
• named tests for every public tool.

Quick start

Install the published alpha package:

python -m pip install "alptugharun-ai-workbench-mcp==0.1.0a2"

Confirm the local package can load before touching host configuration:

alptugharun-ai-workbench-mcp --doctor

A healthy result reports the package/Python versions, all three tool names, catalog counts and the read-only permission boundary. It makes no network request and does not prove host compatibility; it tells you whether the local package layer is healthy.

Then point a stdio-capable MCP host at the server:

Launch command:

alptugharun-ai-workbench-mcp

This repository documents the stdio server itself. For the host we have actually exercised, use the copy/paste Cursor setup and 3-tool verification guide. Other MCP clients can differ, so use their current documentation rather than assuming Cursor's configuration is portable.

Security model

Every public tool declares:

{
  "readOnlyHint": true,
  "destructiveHint": false,
  "idempotentHint": true,
  "openWorldHint": false
}

The implementation does not import HTTP clients, subprocess modules, filesystem-write helpers, browser libraries, or provider SDKs.

That does not mean "trust any MCP server." It means this repository keeps its own boundary narrow and testable.

Verify it yourself

python -m unittest discover -s tests -v
python examples/smoke_client.py

CI runs the package and protocol tests on Linux and Windows.

Package / registry status

PyPI: the project publishes through GitHub OIDC Trusted Publishing and the release workflow signs the wheel with keyless Sigstore. Version 0.1.0a2 adds the local --doctor / --version first-run diagnostics; use the PyPI badge above to confirm the release is available before installing that exact version.

A clean Windows virtual environment installed the exact PyPI version successfully, negotiated MCP protocol 2025-06-18, listed all three tools, completed successful render_prompt and get_assistant calls, and returned a bounded error for an unknown tool.

Official MCP Registry: io.github.alptugharun/ai-workbench-mcp is published and currently reports active in the production registry.

Real-host verification: a maintainer-run Cursor 3.20.21 session invoked list_prompts, render_prompt and get_assistant successfully against the published package. This is host evidence, not an independent third-party endorsement or a universal compatibility claim.

See REGISTRY-PUBLISHING.md, HOST-VERIFICATION.md and the failure-first Troubleshooting ladder.

Contributing

Small, reproducible improvements are welcome. The most useful contributions right now are:

• real MCP host verification;
• protocol edge-case tests;
• clearer failure messages;
• documentation corrections;
• narrowly scoped catalog improvements.

Please read CONTRIBUTING.md before opening a PR. If the server works in a host we have not independently verified yet, a reproducible host report is more valuable than a generic “works for me.”

Origin

This project was extracted from AI Social Media Toolkit so the MCP server can evolve as a focused product instead of being buried inside a larger creator/AI repository.

Built by Alptuğ Harun.

If the server gives you a useful first result, starring the repository helps other MCP users discover it. If it fails, open a reproducible report instead — failures with environment details improve the project faster than vanity metrics.

License

MIT — see LICENSE.