MCP servers as tools
Connect your own MCP server over HTTPS and give its tools to characters, to call mid-call like your own tools.
If you already have an MCP server, a character can use its tools during a call. Register the server's address; we list its tools over MCP, keep a checked copy, and the characters you attach it to call them like any of your tools.
Not our MCP server
This page is about your MCP server giving tools to a character on a call. To manage your characters from an AI assistant through our MCP server, see MCP server.
A dashboard page for MCP servers is coming soon. Today, register them through the API with a secret key.
Register a server
POST /v1/mcp-servers:
Prop
Type
curl https://talk.madein10.com/v1/mcp-servers \
-H "Authorization: Bearer $SECRET_KEY" \
-H "Content-Type: application/json" \
-d '{
"name": "Calendar",
"url": "https://mcp.example.com/mcp",
"auth": { "value": "Bearer your-server-token" },
"filler": "One moment."
}'We list the server's tools at once. A server that can't be listed is still kept, with refresh_error saying why in plain words (unreachable, a 401, not MCP, a version we don't speak): fix it, or refresh.
{
"id": "mcps_…",
"name": "Calendar",
"url": "https://mcp.example.com/mcp",
"auth_header": "Authorization",
"has_auth": true,
"enabled": true,
"timeout_ms": 10000,
"filler": "One moment.",
"protocol_version": "2026-07-28",
"server_info": { "name": "…", "title": null, "version": "…" },
"tools": [
{ "id": "mcp_mcps_…_get_weather", "name": "get_weather", "title": null, "description": "…", "input_schema": { "type": "object" } }
],
"skipped_tools": [{ "name": "bad tool", "reason": "its name isn't one we can use (…)" }],
"refreshed_unix": 1793300000,
"refresh_error": null,
"character_ids": ["chr_…"],
"created_unix": 1793300000,
"updated_unix": 1793300000
}The credential's value is never shown again: has_auth says there is one.
Endpoints
With a secret key. In the dashboard, MCP servers are for owners and admins.
| Method | Path | Body | Answer |
|---|---|---|---|
GET | /v1/mcp-servers | { data: [McpServer] } | |
POST | /v1/mcp-servers | { name, url, auth?, enabled?, timeout_ms?, filler? } | 201 McpServer, its tools listed |
GET | /v1/mcp-servers/{id} | McpServer | |
PATCH | /v1/mcp-servers/{id} | any of the create's fields; auth: null removes the credential, filler: null clears it | McpServer; a new url or auth lists its tools again |
DELETE | /v1/mcp-servers/{id} | { id, deleted: true }; detached from every character | |
POST | /v1/mcp-servers/{id}/refresh | McpServer, its tools listed again; 422 mcp_refresh_failed with why (the tools it had stay) | |
GET | /v1/characters/{id}/mcp-servers | { character_id, mcp_servers: [{ id, name, tools }] } | |
PUT | /v1/characters/{id}/mcp-servers | { servers: [{ id, tools? }] }, in order | as GET |
Tools are listed when you register a server, when its url or credential changes, and when you refresh it.
Attach a server to a character
{
"servers": [
{ "id": "mcps_…", "tools": ["get_weather"] },
{ "id": "mcps_…" }
]
}tools names the ones the character may use. Left out, it takes all of the server's, including ones the server adds later.
The tools we keep
We keep at most 100 of a server's tools, in its order. A tool is left out, and listed in skipped_tools with why (at most 50 listed), when:
- its name isn't letters, digits,
_,.and-(at most 128), or appears twice; - its
inputSchemaisn't a JSON Schema object of at most 16 KB; - it marks arguments with
x-mcp-headerin a way the MCP spec doesn't allow.
Longer descriptions are cut at 1,000 characters. Tool annotations are ignored.
On a call
- The attached servers' kept tools come after your own tools. Names are made safe for the character (MCP's
.and other characters become_, at most 64;_2,_3… when a name is taken). - At most 32 tools a call in all, and 48 KB of MCP schemas and descriptions; the rest are left out.
- When the character calls one, we send
tools/callto your server within itstimeout_ms.params._metacarries"com.madein10/call": { call_id, character_id, tool_call_id, metadata }, so your server knows which call asks. - The character gets
structuredContentwhen there is one, else the content as text (images, audio and files are named, never passed on); an empty result is "Done."; at most 16 KB. - A result with
isError: trueis given to the character in one line ("it answered with an error: …"), as MCP means it for the model. Other failures are told in our words: no answer in time, a status other than 2xx, a redirect, too large, a stream cut off, an MCP error. - A tool that asks for more input mid-call fails ("it asked for more input, which a call can't give"): sampling and elicitation aren't offered.
- Nothing is retried. Each use is logged with your own tools' (never its arguments, answer or error text) and shows in the call's
tool_calls.
Protocol versions
We speak MCP's current revision, 2026-07-28, first, and fall back to the earlier handshake (2025-11-25, 2025-06-18, 2025-03-26) for servers that need it, keeping their session. The version a server spoke is kept in protocol_version. The older HTTP+SSE transport (2024-11-05) isn't spoken.
Every request carries User-Agent: acme-mcp/1 and goes only to public addresses checked when connecting, through no proxy, following no redirects.
Limits
| Servers per organization | 10 (409 asset_limit) |
| Servers per character | 5 |
| Tools kept per server | 100 |
| Tools on a call (yours and MCP's together) | 32 |
timeout_ms | 500 to 15,000 (10,000 by default) |
Credentials
OAuth isn't offered for your servers: give a static credential in auth. Tools are listed again when you refresh or change the address or credential, not on a schedule.