MCP

AI agents reach Deplo at POST /api/mcp, speaking the Model Context Protocol, revision 2026-07-28. This page is the wire-level reference. Setup, connecting an agent, and the full list of tools are on MCP server.

The MCP server is a different framing of this same API, not a second one: each tool runs a GraphQL document in-process as the caller's own principal, so every capability gate, folder grant, token scope and two-factor policy on the GraphQL pages applies identically.

Authentication

An ordinary API token, or the OAuth access token a web client obtains. There is no MCP-specific credential.

claude mcp add --transport http deplo https://<deplo>/api/mcp \
  --header "Authorization: Bearer deplo_xxxxxxxxxxxxxxxxxxxxxxxx" \
  --header "X-Deplo-Team: acme"

A request with no usable token answers 401 with WWW-Authenticate: Bearer resource_metadata="<url>", which is where an OAuth client starts discovery.

Protocol

JSON-RPC over plain HTTP POST, stateless: no session, no initialize handshake to keep alive. Every call runs in one team. The default is chosen by X-Deplo-Team when the agent is connected, and every tool takes an optional team argument (id or slug) to work in another. list_teams names the ones the connection can act in.

tools/list returns only the tools the token can actually call. Destructive tools carry destructiveHint, which is what makes a well-behaved MCP client ask its own user first.

curl https://<deplo>/api/mcp \
  -H "Authorization: Bearer deplo_xxxxxxxxxxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"list_apps","arguments":{}}}'

See also