API, MCP and A2A
Three doors to the same service, with the same rules: the MCP connector for AI assistants, the Kelki Agent for conversation, the API for applications that want to track their requests.
Mapping
| MCP connector | Kelki Agent | API v1 |
|---|---|---|
trouver_service | the need, in conversation | POST /trade-matches |
lister_prestations | the agent's buttons | GET /trades/{trade_id} (+ GET /consent-texts/{version}) |
verifier_lieu | the place question | GET /places |
demander_rappel | consent at the summary step | POST /requests then POST /requests/{request_id}/confirmation |
| The same kelki.ai confirmation page, opened from the “Confirm my request” e-mail. | ||
| — | — | tracking specific to the API: GET /requests, /events, webhooks, /account |
Which to choose
- The MCP connector: you build (or use) an AI assistant and want it to find a professional for the user. No account or key, limited per IP address, no tracking.
- The Kelki Agent through the API (
/agent/conversations): you want a ready-made conversation in your app, with tracking and webhooks (guide). - The API with your own screens: you want to control the flow, without an AI model, with tracking by reference and webhooks (guide).
For another agent to talk to the Kelki Agent (A2A), the principle is the same as through the API: same steps, same consent, same e-mail confirmation.
Shared rules
- Same closed lists and same checks as the form; “Autre besoin” (other need) with a free need of 3 to 160 characters.
- Same exclusions: health trades, trades under review.
- Same e-mail confirmation within 48 hours, then erasure; 3 requests per 24 hours for the same contact, all channels; duplicates ignored.
- Same versioned consent text, recorded word for word; a single professional; origin never shown to the professional.