Every week my Claude and Codex quotas reset, and whatever I haven’t used is gone. For 25 days (to Oct 7), an agent has spent that remainder on 111 tasks for me, many of which I’d never have thought to ask for.
What sold me on it: it took a folder of raw Bakeoff (my startup, which tests coding agents on a company’s own code) experiment results and built Decision Lab, an offline prototype that turns those results into a decision on which coding agent a customer should use. I hadn’t asked for it. I called the first version “an amazing demo” but “overly busy,” and the next run read that comment and rebuilt it. My verdict: “pretty close to something I could show a customer.”
It’s called RemainderBot, and today I’m open-sourcing it. When a reset is a few hours away and there’s quota left, it picks a task off your task lists, issues, and notes, finishes it, and opens a pull request in a private repo for you to review. I think of it as a part-time assistant for the work I’d never get to.
Of the 111 pull requests it has opened, I’ve merged a third, closed a third with a comment the next run reads, and still owe a response on the last third. It produces work faster than I review it:
Things I’d never have thought to ask for
As a first-time founder, I spend my time on Bakeoff’s big items. RemainderBot reads my task list, meeting notes, repos, and calendar, and picks from the rest. Some of what it picked:
- Legwork on startup hypotheses the outside world can answer. A census of coding-agent setup files across 31 public GitHub orgs, which turned up an observability company running CI evals on changes to its AGENTS.md (a practice I’d assumed would be useful but had never seen anyone actually do).
- Prep docs for customer, cofounder, or advisor development calls. A guide for a call with a VP of Product, another for an advisor call with an ex-CPO (“Good guide here”), and one for a co-working day with a cofounder candidate.
- Follow-ups I owed and hadn’t gotten to. Notes to a cofounder candidate, a former colleague, and an executive’s assistant who’d offered to set up a meeting. I sent them with edits (it never sends anything itself).
- A business case for a side venture. A calculator that compares three business cases for selling a healthcare-related product by profit potential, with every assumption labeled and defaulted to something reasonable. My comment on the PR: “This is amazing. It puts hard thinking behind options that [my potential cofounder] and I only discussed at a high level.”
- Blog drafts for ideas I jotted down. It drafted my first Bakeoff process post, and an early version of this one.
Most of what I merged was legwork I’d put off and stakeholder work I wouldn’t have prioritized, so I wrote that into its instructions: “Take work I would delegate, not work I would do myself.”
It runs on tokens you were going to waste
Claude and Codex subscriptions both meter use in a 5-hour window and a weekly one. RemainderBot only wakes up when the weekly reset is less than 4 hours away with at least 10% left, and it only spends quota your plan already includes: no API keys and no extra usage billing. It runs on whatever hardware you’d like - a $5 VPS or a spare Mac Mini. Do take care to give resources to the bot alone.
Last month it used about 300M tokens, roughly 10% of my total token usage. In my best Codex week, it used about 60% of the quota that would otherwise have gone to waste.
A few minutes of your time per PR
After setup, RemainderBot only costs me review time. Each PR says what it built, how it checked it, and the one step to ship it. I merge or close it with a comment, and the next run reads the comment. That’s how I steer it.
Now you can run it too
To run it, clone RemainderBot into a private repo of your own and follow the README. You write a GOALS.md that says what matters to you, then point it at your sources: GitHub issues and repos, Google Tasks and Docs, local Markdown files, or any command that prints Markdown, like an export from Linear or Notion.
Here’s one run, start to finish, for a fictional founder whose GOALS.md says to get her dog-grooming scheduler, Fernhill, to 50 paying salons:
Your next reset is less than 7 days away, but you don’t have to wait for one to try it. Run one small task now (about 20 minutes and a few percent of your weekly quota), then post what it built in this discussion, including what it got wrong. The bad picks are how I’ll make it better.
Thanks to Scott Bailey for running an early version on his machine, and Jackie Hyman for editing.