Run the chat gateway¶
The gateway is the human face of the platform: a web chat where the AI is steered to answer from MCP tool data rather than its own memory.
You need the MCP server running in HTTP mode first.
git clone https://github.com/okfn/mcp-chat-gateway
cd mcp-chat-gateway
uv sync
Configure the AI provider, either by editing local_settings.py or
with environment variables:
| Variable | Example | Meaning |
|---|---|---|
AI_API_KEY |
sk-... |
Key for your LLM provider |
AI_BASE_URL |
https://api.deepseek.com |
Any OpenAI-compatible endpoint |
AI_MODEL |
deepseek-chat |
Model name |
MCP_URL |
http://127.0.0.1:8063 |
Where the MCP server lives |
Then run it:
uv run python app.py # serves on http://127.0.0.1:8064
Open http://127.0.0.1:8064, pick a suggested question or type your
own, and watch the answer arrive with its tables, charts and source
links.
No datasets yet? Ask about the example tools
A freshly cloned server already serves its bundled example tools, so you can test the full loop before installing any real catalog: ask the chat what it can answer, call an example tool through it, and check that the answer comes back with its sources.
Disposable by design
The gateway is intentionally simple: plain HTML/JS/CSS, a minimal Flask backend, no frontend framework. It exists to test and demonstrate the platform; expect it to change quickly.