Kelki in your AI assistant
Ask your AI assistant: "Find me a plumber for a leak under the sink". With the Kelki connector, ChatGPT, Claude or Le Chat prepares your request with you; once you confirm by email, it goes out to local tradespeople across the 51 states, and a professional calls you back.
Your request in 4 simple steps, right in your chat
- 01The tradeyour assistant suggests the trades that match your needs, chosen from the 1,000+ trades available on Kelki.
- 02Your project and timelineyou pick from the same options as the “Call me back” form on kelki.ai (trade services, or “Other request” in a few words), then your timeline. You can add extra details if needed.
- 03Your cityyour assistant verifies your city and ZIP code.
- 04Your contact details and consentname, phone number, and email, a quick summary, and your consent to be called back.
One click in the email, and your request goes out
Right after, you’ll receive an email from Kelki with a summary of your request and a “Confirm my request” button. One click, and your request is sent to trade pros in your area: only one pro calls you back to discuss it. The link stays active for 48 hours; unconfirmed requests are automatically deleted, including contact details.
The service is completely free for you: tradespeople fund Kelki by paying 60 USD excluding tax for the requests they accept. The estimate, appointment and final price are then arranged directly with the professional.
Add Kelki to your assistant
Connector URL: https://kelki.ai/mcp — no account or password required.
- Claude: Customize › Connectors › “Add custom connector”, paste the URL, select “no authentication”.
- ChatGPT: Plugins › “+” › Custom MCP server, paste the URL, without authentication (depending on your ChatGPT plan).
- Le Chat: (Mistral): Connectors › “Add connector” › “Custom MCP connector”, paste the URL.
https://kelki.ai/mcp
Step‑by‑step instructions and setup for Claude Code, Cursor, or VS Code: MCP Connector. On kelki.ai, WebMCP‑compatible browsers also support the “demander_rappel” tool: it fills out the “Call me back” form on the page, which you submit yourself.
The Kelki Agent: finds the pro you need
On every kelki.ai page, the “Kelki Agent” bubble is ready to help: “I have a leak”, “I’m locked out”, “I need a caterer”. It identifies the trade, offers quick‑click options for services, timeline, and city, collects your info, and shows a summary: just check the consent box and submit your request, just like the “Call me back” form.
Your own AI agent can also run the search via the A2A (Agent2Agent) protocol: it asks the same questions, then you receive the “Confirm my request” email. One click, and your request goes out to pros in your area.
Your data
The connector only receives what your assistant forwards for your request: the trade, the selected service (or your need in a few words), the timeline, city and ZIP code, your name, phone, email, any comments, and the name of the assistant used (e.g., “ChatGPT”). Your full conversation stays in your assistant: Kelki only gets these specific fields.
This information is used to send you the confirmation email, and, once you click, to route your request to trade pros in your area. Any request not confirmed within 48 hours is deleted (only trade, service, timeline, and city are kept anonymously, with no contact info). Details, retention periods, and rights: privacy policy.
For developers and platforms
- Full connector documentation (the 4 tools, schemas, JSON‑RPC examples, and setup): MCP Connector.
- Remote MCP server: https://kelki.ai/mcp, Streamable HTTP transport, JSON responses, stateless, no authentication; protocol versions 2026‑07‑28, 2025‑11‑25, 2025‑06‑18, and 2025‑03‑26.
- Tools: trouver_service, lister_prestations, verifier_lieu (read‑only) and demander_rappel (creates a request to confirm by email).
- Resources: kelki://metiers, kelki://accord, kelki://metier/{metier}; prompt trouver_un_professionnel.
- A2A Agent: https://kelki.ai/a2a, JSON‑RPC 2.0 (versions 1.0 and 0.3: SendMessage / message/send, SSE streams, GetTask, CancelTask), no authentication; task in input‑required state at each question, with data payload (step, choices, expected fields, summary, consent copy); after consent, request pending email confirmation (completed task). Agent card: /.well‑known/agent‑card.json.
- Server card: /mcp/server‑card · ARD manifest: /.well‑known/ard.json · overview for assistants: /llms.txt.
Questions?
Email us at contact@kelki.ai: we answer every message. Kelki is published by D.DESIGN & CONSEIL, Paris.