CanDocsOpen Can

Hosted MCP server

Can speaks MCP (Model Context Protocol), so AI agents like Claude and ChatGPT can read your boards, pick up tickets, and work them through your columns. The hosted server requires nothing to install.

Connect in one paste

The hosted MCP server lives at:

https://mcp.can.randomfact.com

Two ways to connect:

  • From Can — open Workspace settings → Connections and use the Connect to Claude or Connect to ChatGPT buttons. They open the client's connector page with the URL pre-filled.
  • From your AI client — add a custom connector / remote MCP server and paste the URL above. Works with claude.ai, ChatGPT (Developer Mode), Cursor, and any client that supports Streamable HTTP with OAuth 2.1.

Either way, a browser tab pops for you to sign in with your Can account. The agent then acts as you: it can only see and change what you can.

In Claude Code, install the Can plugin instead. It bundles this server — same tools, same sign-in — and adds the /can:pr workflow command.

What agents can do

The hosted server exposes a curated tool catalog covering the day-to-day workflow:

  • Read: list_workspaces, get_workspace, list_boards, get_board, list_columns, list_tickets (filter by columnName — "what's in Review?" — or by status: todo, in-progress, done, wont-do, backlog), get_ticket, list_tags, list_comments
  • Find: search_tickets (semantic search), resolve_ticket (find a ticket by key like LAUN-7 with no other context)
  • Write: create_ticket, update_ticket, bulk_status_tickets, create_tag, create_comment, update_comment
  • Guidance: get_help returns workflow documentation written for agents

Destructive structural operations (deleting boards, tags, or comments) are intentionally not exposed on the hosted server. If you need those, use the REST API or the local MCP package.

The agent ticket workflow

The server ships instructions that teach agents a consistent loop:

  1. Read the ticket once, by key (resolve_ticket).
  2. Claim it in a single update_ticket that sets both agentAssigneeId — the ticket card then shows the agent's glyph (Claude, ChatGPT, Copilot, Cursor, and Gemini have icons, and a family variant such as claude-code or gpt-5 gets its family's icon; other agents get an initials avatar) — and the columnId of an in-progress column.
  3. Do the work.
  4. Move it to done, once the work is accepted.

Writes answer with a summary of the ticket rather than the whole thing, so an agent never re-reads a description it just sent. Listing and searching work the same way: each ticket comes back without its description, plus a 160-character preview.

You can watch all of this happen live on the board, and every change the agent makes is recorded in the ticket's activity history labeled with the agent's client name — stored as mcp:<client name>, which board exports keep and the History tab shows as a muted via <client name> after the actor.

Security model

  • Each agent session authenticates with OAuth 2.1; tokens are user-scoped.
  • Your workspace roles apply unchanged — an agent acting for a member cannot do admin-only things.
  • Revoking access (or being removed from a workspace) takes effect on the agent's next call.