Connect your assistant
Add our MCP server to Claude Code, Claude Desktop or Cursor, and set up characters and website widgets by asking.
Our MCP server lets your own AI assistant do the setting up for you. Connect it with one of your secret keys, then say what you want: "put a receptionist on my site that greets in Persian and English". The assistant makes the character, puts it on your site, writes the widget's words in each language (translating them itself), and hands over the tag, or pastes it into your site if it can edit it.
It works through the same checks as the API: every tool calls /v1 for you, with your key's permissions, limits and env.
Names that will change
The product has no name yet. The server's name, acme, and its address, talk.madein10.com/mcp, are placeholders that work today; they will change when the product has its name.
The address
https://talk.madein10.com/mcpSend a secret key as Authorization: Bearer sk_live_…. Use a test key (sk_test_…) to try things on test deployments first: a test key sees and makes test deployments, test calls and test webhooks only; a live key, live ones. Characters and knowledge belong to your organization and are shared by both. Make keys on the dashboard's API keys page.
Add it to your client
One command:
claude mcp add --transport http acme https://talk.madein10.com/mcp --header "Authorization: Bearer sk_live_…"Any client that speaks MCP's Streamable HTTP transport and can send a header connects the same way.
ChatGPT's connectors and Claude's remote connectors (Claude.ai, and Claude Desktop's Settings → Connectors) want OAuth, not a key in a header. Signing in with OAuth, picking your organization and live or test, is planned. Until then, use a client above.
Check the connection
The MCP Inspector lists the tools from the command line:
pnpm dlx @modelcontextprotocol/inspector@2.13.0 --cli https://talk.madein10.com/mcp \
--transport http --header "Authorization: Bearer sk_test_…" --method tools/listRules it keeps
- The key is read exactly as
/v1reads it. No key, a wrong or revoked one, or a publishable key is 401invalid_api_key. Revoking the key stops the next request. - One budget with
/v1. Your key's 600 requests a minute count every MCP request, and each tool also counts the/v1requests it makes (most make one; some, likeset_guardrailsoradd_knowledge, up to six). Past it: 429rate_limitedwithRetry-After: 60. See Limits. - Nothing destructive without a yes. Tools that delete, revoke or roll a key do nothing until the assistant asks you and calls again with
confirm: true(Tools). - No secret shown twice. A deployment's key is shown when it's made or rolled, a webhook's signing secret when it's made, as on the dashboard and
/v1. - No live calls. No tool starts a call: a call needs a browser and a microphone.
- Not for browsers. A request from a web page (an
Originother than ours) is refused with 403, and there is no CORS. A secret key never belongs in a web page; browser-based clients go through their own server.
What it speaks
For the curious and for client authors:
POST /mcp, one JSON-RPC message per request, always answered as one JSON object (never a stream).GETandDELETEare 405: there's no stream to open and no session to end.- The current MCP revision, 2026-07-28 (no handshake;
server/discover), and the earlier ones from 2025-03-26 to 2025-11-25 throughinitialize. No session id is kept: every request stands alone. - Tools, resources and prompts. No sampling, elicitation, roots, subscriptions, logging or completions.
- A request body of at most 4 MB (413 above that).