MCP: use OneType from Claude Code, Cursor and other AI tools

OneType is an MCP server (Model Context Protocol, Streamable HTTP). Any MCP client can call it with an API key: Claude Code, Cursor, Claude Desktop, or your own agents. There are two servers:

Server	URL	For	Tools
Workspace	`https://api.onetype.ai/v1/mcp`	Your team and your own automations: read agents and conversations, search knowledge, ask an agent, and build agents	Read tools: `list_agents`, `get_agent`, `list_conversations`, `get_conversation`, `list_messages`, `search_knowledge`, `ask_agent`. Builder tools: see Builder tools
One agent's docs (hosted)	`https://api.onetype.ai/v1/mcp/agents/YOUR_AGENT_ID`	Your developers and customers: the same cited answers as your chat, inside their editor	`search_docs`, `ask_agent`

Status: both servers are Available. Both need Premium: on Free, MCP connections are refused with 422 PLAN_LIMIT.

1. Create a key

Workspace server: in Developer › API keys, create a restricted key (rk_live_…) and tick only the permissions you need:
- agents.read for list_agents and get_agent
- conversations.read for list_conversations, get_conversation and list_messages
- search.debug for search_knowledge
- conversations.write for ask_agent
- for a builder tool, the permission of the same action in the app (for example agents.update for update_agent, knowledge.write for add_website)
One agent's docs: on the MCP page, open the agent's card and create its key. The key holds mcp.agent and is limited to that one agent, so it can't read anything else in your workspace. You can safely share it with the developers who use your docs.

Keys are shown once. Revoking one takes effect on the next request.

2. Connect your client

Claude Code

claude mcp add --transport http onetype https://api.onetype.ai/v1/mcp \
  --header "Authorization: Bearer rk_live_…"

One agent's docs:

claude mcp add --transport http onetype-docs https://api.onetype.ai/v1/mcp/agents/YOUR_AGENT_ID \
  --header "Authorization: Bearer rk_live_…"

Cursor (.cursor/mcp.json), Claude Desktop and other clients that take JSON:

{
  "mcpServers": {
    "onetype-docs": {
      "url": "https://api.onetype.ai/v1/mcp/agents/YOUR_AGENT_ID",
      "headers": { "Authorization": "Bearer rk_live_…" }
    }
  }
}

Developer › MCP shows both snippets with your URL filled in.

Then ask in your editor: "Using onetype-docs, how do I verify a webhook signature?" The answer comes with its sources.

What each tool does

Tool	Server	Returns	Billing
`list_agents`, `get_agent`	workspace	Agents, as the REST API returns them	free
`list_conversations`, `get_conversation`, `list_messages`	workspace	Conversations and their messages	free
`search_knowledge`	workspace	The best-matching passages from an agent's knowledge, with sources	free
`search_docs`	one agent	The best-matching passages from that agent's knowledge, with sources	free
`ask_agent`	both	A full answer from the agent, with citations	one answer, like a chat reply

Passages, messages and answers are marked as untrusted content for the calling model, so text inside your documents can't instruct the client's AI: on both servers, every field your documents, your visitors or a model wrote (a passage's header and text, a message's text, a source's title, an agent's name and instructions) comes as an object { "untrusted_text": "…" }, never a bare string. Read the string from untrusted_text. Every call is recorded in your audit log (mcp.tool.called). A tool result never starts anything on the server.

Builder tools

The workspace server also builds and runs agents. Each tool works like the same action in the app: the same permissions, plan limits and audit log.

Group	Tools
Account	`get_account`
Agents	`list_models`, `create_agent`, `update_agent`, `publish_agent`, `take_agent_offline`, `resume_agent`, `delete_agent`
Knowledge	`list_sources`, `get_source`, `add_website`, `add_pages`, `add_text`, `add_file`, `resync_source`, `remove_source`, `list_qna`, `add_qna`, `update_qna`, `delete_qna`
Chat and site	`set_allowed_domains`, `get_look`, `update_look`, `apply_look`, `set_bubble`, `get_install_snippet`, `test_agent` (billed like a chat reply)
Templates and links	`list_templates`, `get_template`, `get_builder_prompt`, `get_checkout_link`, `get_connect_link`, `get_app_link`

Destructive calls wait for a person. Deleting an agent, a source or a Q&A, publishing an agent or taking it offline, and on a live agent publishing its look, changing its allowed domains or changing what visitors are told or what is collected from them, answer needs_confirm with a link. A person signed in to OneType approves it in the browser; the same call with confirm set to the given id then goes through once, within 10 minutes. Payment and connecting other tools also happen in the browser, through the links the tools return. Changes are limited to 60 a minute, and create_agent to 10 an hour.

For developers of other MCP clients

POST takes one JSON-RPC message. initialize answers with an Mcp-Session-Id header; send it on later requests.
Supported protocol versions are 2025-11-25 and 2025-06-18.
Methods: ping, tools/list, tools/call. Notifications answer 202. DELETE ends a session, and GET answers 405. Sessions end after a day unused.
tools/list offers a tool only when the key holds its permission.

Errors

Response	Meaning	Fix
`401`	Missing or revoked key	Check the `Authorization` header
`403 ORIGIN_NOT_ALLOWED`	Called from a browser page	Call it from a server or a desktop client. Keys never belong in a browser
`403 AGENT_SCOPED_KEY`	An agent's key used on the workspace server	Use the agent's URL, or a workspace key
`403 AGENT_OUT_OF_SCOPE`, `422 PLAN_LIMIT`	The key isn't for this agent, or the workspace is on Free	Use that agent's key, or upgrade to Premium
`400 SESSION_REQUIRED`, `PROTOCOL_VERSION_MISMATCH` · `404 SESSION_NOT_FOUND`	Session handling	Re-run `initialize`
JSON-RPC `-32601` / `-32602`	Unknown method, or unknown tool or bad arguments	Call `tools/list` for what this key may use
A result with `isError: true`	The tool was refused (e.g. a missing permission)	Its code and message say why

People can also connect with OAuth instead of a key: claude.ai and Claude Desktop (a custom connector) and ChatGPT sign in through OneType's consent screen. The MCP page in the app has the steps for each tool.