Give an AI agent live African FX rates with MCP

Search results give a model pages to read, most of which show a mid nobody transacts at and few of which carry a timestamp the model can trust. A tool gives it structured, dated data with definitions attached. Modan's MCP server returns the same objects as the REST API — provider entries with rate, fee, spread_bps, vs_mid_bps, rate_type and last_updated — and the server's own instructions tell the model what each field means, so it does not have to guess that spread is measured within a rate type.

• get_rates — the full provider book for a corridor, with the independent mid.
• convert — what an amount delivers per provider net of fees, and who delivers the most.
• fetch_rates — mid-market rates from one base to up to 20 quotes in one call.
• get_history — timestamped observations over a window, optionally one provider.
• list_corridors, list_providers, list_currencies — discovery, so the model can answer "what do you cover?" without browsing.

The server is remote — nothing to install — at https://modan.io/api/mcp, streamable HTTP with plain JSON responses, authenticated with an X-API-Key header. The MCP setup page has the exact configuration for Claude Code, Cursor, VS Code and a generic JSON block for other clients, plus a keyless tools/list probe to confirm the connection before adding a key.

claude mcp add --transport http modan https://modan.io/api/mcp \
  --header "X-API-Key: mdn_live_YOUR_KEY_HERE"

Ask "Which provider gives the most naira for £500 right now?" and the model calls convert with GBP, NGN and 500, then answers with the provider, the net amount, and — if it is well behaved — the observation time and the plan's freshness. Ask "how has LemFi's GBP→NGN rate moved this month?" and it calls get_history with a provider filter. Ask "what corridors do you cover into GHS?" and it calls list_corridors and filters.

• Each tool call uses one request of the key's daily quota — 50 on a free key, 250 on Individual, 1,000 on Team — shared with the REST API.
• Answers are labelled hourly or real-time by plan; the model receives data_freshness and as_of and can repeat them. Ask it to.
• Reference prints (official, parallel) come back labelled and are never chosen as best; stablecoin corridors carry no mid.
• ChatGPT's connectors cannot send an API-key header today; a Custom GPT Action built from openapi.json can call the REST endpoints instead.