WorkTracker › Features › Agent access (MCP)

Your AI agent can use WorkTracker too

A built-in MCP server lets Claude, ChatGPT, Gemini and any other MCP-capable agent read your timeline, check your tracked time and manage your todos — with a personal, revocable token.

One endpoint, any agent

WorkTracker speaks the Model Context Protocol — the open standard AI assistants use to reach external tools — over plain HTTP. Point your agent at one URL and it can answer questions like “what did I work on Tuesday?”, “how much time did I track on ACME this month?” or “add a todo to follow up the review”.

Endpoint: https://time.itcm.se/api/mcp

Add it in one click

WorkTracker is a full OAuth server, so most apps connect with no config file and no token to copy. In Claude, ChatGPT or any client with a “custom connector” or “add MCP server” box, just paste the endpoint URL. The app auto-discovers the rest, opens a WorkTracker sign-in, shows you exactly what it will be allowed to do, and you click Allow. That’s it — the app registers itself and gets its own credentials behind the scenes.

Access is yours to revoke: every connected app is listed under Profile → Agent access, one click to disconnect.

What your agent can do — and what it cannot

• Read any day: every collected event with source, project, duration and its AI summary.
• Summarize tracked time per project for a day or a date range.
• See the company project vocabulary (projects with their clients).
• List and add todos, move them through their lifecycle (new → in_progress → blocked → completed) and append run-log notes with links to pull requests and builds.
• It cannot create or edit tracked time, push to Toggl, or see items you hid — the write surface is deliberately todos-only, so no agent can fabricate billable hours.

AI tasks — you decide what the agent works on

A todo only becomes fair game for an agent when you flag it an “AI task” on the Todos page. Your agent can list exactly those (and only those), so you can tell it “work through my AI tasks” and it leaves the rest of your list alone. An agent can complete an AI task, but never one of your own todos.

Enrich each AI task with instructions: how to complete it and how to verify it is done. The agent reads them over MCP before it starts — “update the pricing table, run the tests, then reply with the diff” — so the task carries its own definition of done.

As it works, the agent moves the task through its status and writes notes to the run log — progress, decisions and links to the pull request it opened — which you read right on the todo in WorkTracker. Want this running on a schedule? That’s /features/wt-worker.

Connect Claude

In Claude Desktop (or claude.ai), open Settings → Connectors → Add custom connector, paste the endpoint URL, and sign in when prompted — the OAuth flow does the rest.

Prefer a config file, or using the Claude Code CLI? Create a token under Profile → Agent access and pass it as a header instead:

claude mcp add --transport http worktracker \
  https://time.itcm.se/api/mcp \
  --header "Authorization: Bearer wtk_…"

Connect ChatGPT

In ChatGPT enable developer mode under Settings → Connectors, add a connector with the endpoint URL, and complete the WorkTracker sign-in. ChatGPT discovers the tools automatically.

Connect Gemini

For the Gemini CLI, add the server to ~/.gemini/settings.json:

{
  "mcpServers": {
    "worktracker": {
      "httpUrl": "https://time.itcm.se/api/mcp",
      "headers": { "Authorization": "Bearer wtk_…" }
    }
  }
}

Connect anything else

Any client that implements MCP over the streamable HTTP transport works the same way: it either signs in through the built-in OAuth flow or sends a personal token as a bearer header. No SDK and no webhook setup — one URL.

Secure by construction

• Access is personal: an agent — whether it signed in with OAuth or uses a token — sees exactly what its owner sees, never other members’ data.
• Only hashes of tokens and secrets are ever stored, sign-in uses PKCE, and every connected app or token is revocable with one click under Profile → Agent access.
• Hidden items never reach an agent, and privacy-filtered content was already removed before any agent could ask — the same guarantees as at /privacy.