TAAFT MCP server

Remote, stateless Model Context Protocol server for finding AI tools in the TAAFT directory.

It exposes:

• search_tools — search by query, type, sort, and result limit.
• get_tool — retrieve public details by the slug returned from search.

The MCP endpoint is POST /mcp. GET /healthz is an unauthenticated liveness endpoint.

Install

The public endpoint is:

https://theresanaiforthat.com/mcp-server/

It uses Streamable HTTP and requires no authentication.

Cursor

The repository root is a Cursor plugin. For local plugin development:

mkdir -p ~/.cursor/plugins/local
ln -s "$(pwd)" ~/.cursor/plugins/local/taaft

Run Developer: Reload Window, then confirm taaft appears under Customize. Remove or disable a manually configured taaft server to avoid a duplicate name.

Claude

Add the connector directly to Claude Code:

claude mcp add --transport http taaft \
  https://theresanaiforthat.com/mcp-server/

Or install the Claude plugin from this repository:

claude plugin marketplace add taaft/mcp
claude plugin install taaft@taaft-plugins

For Claude.ai and Claude Desktop, add a custom connector under Customize → Connectors using the public endpoint and authentication type None.

Codex

Add the server directly to Codex CLI or the Codex IDE extension:

codex mcp add taaft \
  --url https://theresanaiforthat.com/mcp-server/

For plugin-capable Codex surfaces:

codex plugin marketplace add taaft/mcp
codex plugin add taaft@taaft-plugins

You can also install it interactively from /plugins. Codex IDE supports direct MCP configuration but does not currently install plugins.

ChatGPT

Enable Developer mode, open Plugins, select +, and register:

• URL: https://theresanaiforthat.com/mcp-server/
• Authentication: None
• Transport: Streamable HTTP

The repository contains the Codex/OpenAI plugin package, but a ChatGPT plugin also needs an .app.json mapping to the plugin_asdk_app... identifier that ChatGPT creates when the endpoint is registered. That account-specific file cannot be generated before registration.

Visual Studio Code

This repository includes .vscode/mcp.json. Opening the repository in VS Code loads the server at workspace scope. To install it globally without cloning:

code --add-mcp \
  '{"name":"taaft","type":"http","url":"https://theresanaiforthat.com/mcp-server/"}'

A VS Code extension is unnecessary for a fixed remote endpoint. Public @mcp gallery inclusion is a separate, curated GitHub MCP Registry process.

Google Antigravity

After cloning the repository:

mkdir -p ~/.gemini/config/plugins
cp -R plugins/antigravity/taaft ~/.gemini/config/plugins/taaft

The plugin uses serverUrl, the current Antigravity key for remote MCP servers. No transport or authentication fields are needed. Restart Antigravity after copying the plugin.

Prerequisites

• Node.js 22+
• Access to the internal TAAFT HTTP API

Local development

npm install
cp .env.example .env

Set at least TAAFT_API_BASE_URL in .env, then run:

npm run dev

By default, the service listens on http://localhost:3000/mcp.

The upstream adapter expects:

• GET {TAAFT_SEARCH_PATH}?query=...&type=...&sort=...&limit=...
• GET {TAAFT_TOOL_PATH_TEMPLATE}, replacing {slug} with the URL-encoded slug

Search may return an array or an object containing tools, results, or data. Detail may return the tool directly or under tool or data. Change the adapter in src/taaft-client.ts if the internal API contract differs; do not add more environment-driven response mapping.

Example client configuration:

{
  "mcpServers": {
    "taaft": {
      "url": "https://theresanaiforthat.com/mcp-server/"
    }
  }
}

Production

Set:

• NODE_ENV=production
• PUBLIC_BASE_URL to the external HTTPS origin
• TAAFT_API_BASE_URL and, if required, TAAFT_API_KEY

Build and run:

docker build -t taaft-mcp .
docker run --rm -p 3000:3000 --env-file .env taaft-mcp

Terminate TLS at your ingress or load balancer. The application is stateless, so it can run multiple replicas without sticky sessions. Apply rate limits at the ingress because the MCP endpoint is intentionally public; application-level rate limiting is ineffective once replicas are added.

Verification

npm run typecheck
npm test
npm run build

Then connect with the MCP Inspector:

npx @modelcontextprotocol/inspector http://localhost:3000/mcp