TokenLab OpenAI Apps Model Explorer

Lightweight TokenLab Model Explorer prototype for ChatGPT/OpenAI Apps SDK. It exposes an MCP Streamable HTTP endpoint plus an interactive MCP Apps widget for:

• browsing TokenLab models from /v1/models, preserving the native decision category

• comparing platform prices, context windows and capabilities from the live /v1/models catalog (data[].tokenlab.pricing, with currency)

• generating contract-checked examples for Chat Completions, Responses, Anthropic Messages, Gemini generateContent, and Jev / System One decisions

This repository is intentionally small so it can be used as a public discoverability asset and a starting point for a hosted ChatGPT app.

Live MCP endpoint: https://exposure-tokenlab-openai-apps-model.vercel.app/mcp

The maintained Vercel project deploys this repository to that hostname. The historical tokenlab-model-explorer.vercel.app deployment is not attached to this project and still serves the older contract; use the maintained endpoint for Jev decision examples.

Run locally

npm install
npm test
npm start

The server listens on http://localhost:8000.

• MCP endpoint: http://localhost:8000/mcp

• Standalone model directory: http://localhost:8000/widget

Opening /widget directly loads the complete public model index, with 50 results per page. Search and category filters run across the entire loaded catalog, not just the current page. The embedded MCP App resource remains a tool-result view and displays no placeholder models before the host supplies a result. MCP discovery tools return every model that matches the category or query, so a category count equals the live catalog's count.

For ChatGPT local testing, expose the MCP endpoint with a tunnel such as ngrok, then add the /mcp URL as a connector in ChatGPT developer mode.

The Express app is also exported as the default module, so Vercel can deploy this repository without a custom adapter.

Related MCP server: APIAny MCP

Environment

TOKENLAB_API_BASE=https://api.tokenlab.sh
PORT=8000
# Optional: server-side catalog cache in ms (default 300000; capped at 600000 = 10 minutes)
CATALOG_CACHE_MS=300000
# Optional: OpenAI Apps domain verification token served at /.well-known/openai-apps-challenge
OPENAI_APPS_CHALLENGE_TOKEN=

Live catalog and freshness

All three tools read GET /v1/models (public, no key) and keep one server-side copy for at most 10 minutes. Every result carries catalog_fetched_at, and every model carries price_fetched_at and pricing.currency. If the catalog cannot be read and no copy younger than 10 minutes exists, the tool returns an explicit error instead of an old snapshot. Tool output contains no top-up or purchase links.

OpenAI Apps domain verification

GET /.well-known/openai-apps-challenge returns the token as plain text. It is taken from OPENAI_APPS_CHALLENGE_TOKEN, otherwise from config/openai-apps-challenge.txt (the token is public, so the file may be committed); with neither, the route answers 404.

Data handling

The server keeps no database and no user data. It does not log tool arguments or requests itself (the only console output is the startup line), sets no cookies, and has no third-party analytics. It forwards no credentials to TokenLab. The only data held is the in-memory public catalog copy (up to 10 minutes). The hosting platform (Vercel) keeps its own request logs under its own retention.

Tools

Discovery inputs

• https://api.tokenlab.sh/v1/models and /v1/models/{id}

• https://api.tokenlab.sh/integrations.json

• https://tokenlab.sh/docs/openapi.json

Notes

• The standalone /widget and Hugging Face static Space share huggingface/index.html. Upload that exact file to the existing kiln4758/tokenlab-model-explorer Space after validation; do not independently edit its pagination or endpoint logic in two places.

• /widget reads the catalog through same-origin /public/v1/models, and /public/v1/models/:model (/public/pricing.json is still proxied for older pages). These read-only routes forward neither credentials nor arbitrary URLs; the static Space retains direct public API reads from its allowed origin.

• This app does not require a TokenLab API key for discovery tools.

• Inference tools should use @tokenlabai/mcp-server; this repo focuses on model exploration and endpoint examples. System One examples are synchronous typed decisions, not chat replies or authorization to execute actions.

• The widget follows the MCP Apps pattern: tools are registered with UI metadata, and the HTML resource is served as an MCP app resource.