Work from Your Own AI (MCP)
Besides the agents you configure inside the workspace, you can connect your own AI assistant — Claude, ChatGPT, Codex, or any MCP-capable client — and have it operate the workspace directly. This is the Model Context Protocol (MCP): an open standard that exposes XPT Tracker as a set of tools your assistant can call.
The point is full, hands-on operation of XPT Tracker from where the work happens — your editor, your terminal. Instead of switching to the web app to file a task, the assistant you're already talking to files it. Instead of copy-pasting a plan into a doc, it writes it straight into the knowledge base. A connected assistant can:
- plan a piece of work and create the project, milestones, and tasks for it;
- pick up a task, read its context, do the work, and move it across the board — following your workspace's AI Workflows;
- write results, decisions, and learnings into the knowledge base;
- take part in channels, brainstorms, and task threads.
How a connection is made
Connecting takes one thing: your workspace's URL, of the form
https://mcp.xpt.ai/<your-workspace>/mcp
Copy it from Settings → Connect to Claude or Settings → Connect to OpenAI, give it to the assistant, and it does the rest — it opens your browser, you sign in and confirm on a short page, and the connection is live. Nothing secret is copied by hand.
The URL names the workspace, so an assistant connected with it can reach that workspace and no other. That is what lets one machine work with several workspaces at once.
Wherever you are asked to name the connection, xpt_<your-workspace> is a good name: it stays readable once there is more than one, and it survives what clients do to names. Avoid spaces — the name becomes the tool prefix you read on every call, and a client rewrites whatever you type into something it can use, so XPT - Acme Inc. comes back as xpt_-_acme-inc.
Set one up
- Connect Claude — a connector in your account, or a server in one project.
- Connect ChatGPT and Codex — the plugin form in the app, and the Codex CLI.
- Connect any MCP client — everything else, and the token path for scripts that cannot open a browser.
- Workflows in Your AI Client — how your workspace's AI Workflows appear in the client as tools and slash commands.
What the assistant can do — the toolset
Tools are named domain_action, so they group by area. At a high level:
| Area | What the tools cover |
|---|---|
| Workspace & people | Who am I, workspace overview, member list |
| Teams | Teams and their states, labels, priorities, members |
| Tasks | List, get, create, update, batch-update, archive, and comment on tasks |
| Projects | List, get, create, update, archive; states, status updates, dependencies |
| Milestones | List, get, create, update, archive |
| Knowledge | Browse the tree, read, search, create, update, delete pages |
| Channels | List, get, create, archive, and post (mentions can spawn agents) |
| Brainstorms | List, get, search, create, and post |
| AI Workflows | Your workspace's processes as runnable tools (wf_task_plan, …) — plus reading and authoring the workflows themselves. See Workflows in Your AI Client. |
So a connected assistant can run essentially the whole workspace, not just read it — and it follows your workspace's processes while doing so.
Safety
- Every tool call goes through the same permission checks as the web app — the assistant only sees and changes what you can.
- A connection stops working the moment you lose access to the workspace, without waiting for anything to expire.
- Mutating tools use an explicit "changes" model, so the assistant can't blank a field by leaving it out.
- Actions appear in the relevant Activity feed, attributed to you.
- Disconnect an app, or revoke a token, at any time from Settings → API tokens and connected apps.
- Insights → Connected clients shows what is talking to the workspace right now, and whether it connected through the browser or with a token.
A connection is different from an agent's provider key (OpenAI/DeepSeek/…). The connection lets an external assistant work inside the workspace; the agent key lets an in-workspace agent call an LLM.