Running a lab means holding a dozen threads at once: who is blocked, which deadline is closest, what shipped last week, and which paper everyone should have read by Friday. The Lab Assistant keeps those threads in one place.

What it does

• Standups from real activity. Connect your GitHub repos and the assistant writes a daily digest from actual pull requests and commits — what shipped, what's in progress, and which PRs have gone stale.
• A shared task board. Assign work, set priorities and due dates, and let every member see their own plate under My work.
• Deadline radar. Track your target venues and get a warning as submission dates approach, with CORE tiers attached.
• Paper watch. Follow a topic and the assistant surfaces the newest relevant arXiv papers as a reading list.

Grounded, not guessed

The assistant answers from your lab's real data through a set of tools rather than free-associating. Ask "what did the team ship this week?" and it reads your repositories. Ask it to create a task and it confirms exactly what it did.

If the data isn't there, it says so. That matters more than sounding confident.

Getting started

Create a lab, share the invite link with your group, and connect a repository. The first standup takes about a minute to generate.

Why grounding matters more than fluency

A lab assistant that invents a deadline is worse than no assistant. The failure is not embarrassing, it is expensive — somebody plans around it.

So every answer is derived from a tool call against real data: your repositories, your task board, your tracked venues. When the data does not support an answer, the correct output is to say so. That constraint costs fluency and buys trust, which is the right trade for anything a team relies on.

The standup problem

Written standups decay. People forget what they did, describe intentions as progress, and the meeting becomes a status performance.

Deriving the digest from actual repository activity changes what is being reported. Merged pull requests, commits, and stalled branches are facts. The assistant's job is to summarise and highlight what has not moved — the stale PR nobody has reviewed for six days is the useful output, not the list of what shipped.

What it deliberately does not do

Assign work. Judge performance. Escalate.

Those are management decisions, and a tool that makes them badly does far more damage than one that declines. The assistant surfaces; a human decides.

Where it fits a research group

The recurring cost in a lab is coordination, not thinking. Who is blocked, which deadline is closest, what everyone should have read by Friday. Those are cheap questions that are expensive to answer because the information is scattered across a repository, a calendar, a chat log and somebody's memory.

Putting them in one place does not make anyone smarter. It removes an hour a week of asking.