time-tracker-mcp

A local MCP server that lets Claude Code read your logged hours from our internal time tracker.

Phase 0 of bench project BP-006 — Remote MCP Skill (remote-mcp-skill). This is the reference implementation: get the tool logic right locally, then port it to AWS.

Why

Open Google Calendar, read last week's meetings, and transcribe them into the tracker by hand. Claude can already see your calendar via the Google Calendar connector — it just can't see the tracker. This closes that gap:

You: What meetings from last week haven't I logged yet?

Claude: [reads calendar, reads tracker, diffs]
        Logged 4h. Calendar shows 7h 45m. Missing:
        - Thu 30 Jul, Prism Standup, 15m
        - Fri 31 Jul, Masterclass, 3h
        Flag: Fri Masterclass overlaps Mentors x Interns 16:30-17:00.

Read-only. It tells you what to enter; it does not enter it.

Related MCP server: Beebole MCP

Tools

Any range works — a day, a week, a month. Totals and grouping are left to the caller: the tools return rows, Claude does the arithmetic.

Install

git clone https://github.com/mikka-oldst/time-tracker-mcp.git
cd time-tracker-mcp
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
cp .env.example .env          # then fill in — see Auth below
.venv/bin/python server.py --smoke

--smoke checks the date conversions offline, then makes one live call per tool. Run it before wiring anything up; it fails loudly and specifically.

Then register it (absolute paths — Claude spawns this as a subprocess with no shell context):

claude mcp add -s user time-tracker -- \
  /full/path/to/time-tracker-mcp/.venv/bin/python \
  /full/path/to/time-tracker-mcp/server.py

Restart your session, confirm with /mcp, and ask it something.

Auth

Two cookie values, pasted once. The server exchanges them for a fresh Convex JWT on demand, so you never hand-paste a token.

DevTools → Application → Cookies → the tracker's domain:

The Clerk session id is read from the cookie's sid claim, so there is nothing else to copy. Leave CLERK_JWT blank — if it is set, it overrides refresh.

Tokens come from Clerk's convex JWT template and last an hour. The default template mints 60-second tokens that expire before you can use them, which is why the template name matters.

How it works, and what surprised us

The tracker is Next.js on Vercel with Clerk auth and a Convex backend. It has no REST API — the browser talks to Convex over a WebSocket. Probing /api/* returns 404 for everything, and the Network tab looks empty unless you filter to WS and read the frames.

Function names and argument shapes came from the ModifyQuerySet frames in that WebSocket. timeEntries:listByWeek takes an arbitrary range despite its name — the UI calls it with a single day.

Three things worth knowing if you extend this:

• Epoch milliseconds, Manila midnight. Not UTC. Getting it wrong shifts entries by a day. Asserted against real captured frames in smoke().

• Convex returns HTTP 200 on function failure. Check status in the response body, not the status code.

• Entries carry 21 fields, most of it Clockify/ClickUp sync plumbing. They get trimmed to 7 before reaching Claude; passing the rest floods the context.

We use httpx against Convex's HTTP query endpoint rather than the official convex Python client, which hung indefinitely with no timeout option.

Limitations

• Local only. The cookie-refresh flow uses Clerk's undocumented client API and authenticates as you. It cannot run on Lambda. A remote deployment needs a Convex deploy key or a Clockify API key.

• Read-only. No writing entries. A human should confirm hours.

• Calendar shows invitations, not attendance. RSVPs are often needsAction, so a meeting on the calendar is not proof you attended it.

• Overlapping meetings need a human rule. The tools surface the conflict rather than guessing which one wins.