GeneralStaff. GitHub Join early access

generalstaff.io · open engine, pre-launch desk

One desk for every AI subscription you already pay for.

GeneralStaff runs AI coding work across the subscriptions you already have. A separate reviewer checks each change before it lands. Co-seats keep the work moving while you are away, and anything only you can decide waits on your desk.

View on GitHub Join early access Watch it work

Open source · AGPL-3.0-or-later · v0.17.0 · bring your own keys

A short film

The desk, running.

The film is coming. Until then, the demos below play in the page.

Loads from youtube-nocookie.com only after you press play.

01 · What it does today

You pay for the models.
The desk keeps them working.

This is version 0.17.0, released 1 November 2026. Each line below comes from the public repository, and you can read the code behind it.

  1. Co-seats

    A seat that keeps working when you do not.

    A co-seat wakes one short agent turn when a cycle finishes, an Aide writes to its inbox, or a timer runs out. The turn reads what happened and starts the next gated cycle. The driver that decides whether a turn runs has no model in it. Usage meters, stops and the deadline live in that driver, not in the prompt.

  2. Runners

    Many accounts, first one under its limit.

    Give a seat an ordered list of runners, each tied to its own account and usage meter. It takes the first runner under its limit and falls back to the next. A meter that fails reads as full, so a broken meter grants no turn. You can stop or start a seat from a phone or another machine.

  3. Review

    One independent review before anything lands.

    Each cycle runs your verification command, checks the diff is not empty, and asks a separate reviewer whether the diff matches the task and leaves your protected files alone. Fail a step and the cycle rolls back. Since v0.17.0, a review that did not finish cannot count as a pass.

  4. Needs you

    Your questions wait in one place.

    A co-seat makes no taste decisions. It parks them, up to a daily cap you set, where generalstaff forks and the dashboard list them. The workflow guide puts each decision to you as one concrete question with a few options and a recommended answer.

  5. Aides

    Named agents with their own memory.

    An Aide is a persistent agent with its own memory, credentials and delegated authority. It can run on another machine. GeneralStaff reaches it through one registered door and writes each exchange as a row in the Aide’s inbox, which gives you an audit trail.

02 · Watch it work

Two flows, played in place.

Each demo runs in the page and starts when you scroll to it. The projects and numbers are invented. The commands and the order of events follow the co-seat guide.

Co-seat turn

Outside work finishes. A seat wakes.

The driver sees an outside change and starts one turn. The turn verifies the outside change, parks a question for you, writes its state file and ends.

Workflow guide

One parked item at a time.

In a Claude Code session the workflow guide puts each parked decision to you as one question, with a few options and a recommended answer.

03 · The verification gate

Nothing merges on a model’s say-so.

The gate is code, not a prompt. A cycle works on a separate bot/work branch, and a change leaves it only after your tests pass and the reviewer signs off.

One cycle

Queued, run, gated, reviewed, merged.

Every step lands in the progress log. Merging into your branch stays off until you set auto_merge.

  • The tests

    Your own verification command runs on every cycle. A non-zero exit fails it.

  • The diff

    An empty diff reported as done fails. So does a diff outside the task’s scope.

  • The reviewer

    A separate reviewer confirms the diff matches the task and respects your protected files.

  • The record

    Each prompt, response, tool call and diff is written to disk for you to read later.

What the gate does not catch

  • Tests that pass for the wrong reason. Scope-match is not correctness, so the gate ratifies them.
  • Failures before a diff exists. A missing toolchain or a crashed engineer leaves nothing to verify.
  • Remote effects of your tools. GeneralStaff does not restrict what the commands you configure reach.

04 · Your subscriptions, your machine

The desk keeps the books.
The accounts stay yours.

GeneralStaff runs on your machine. You bring the keys or the logins, and your provider bills you.

  • Runs on your machine, with a dashboard on 127.0.0.1
  • No telemetry and no hosted tier
  • No GeneralStaff platform fee
  • Co-seat config names key variables and holds no key values

Providers it can use

  • Anthropic API key
  • Claude Pro or Max login
  • OpenRouter
  • Ollama, for local models
  • Aider, and changes from any agent through verify-only

05 · Open source

The engine is open.
Read it, run it, change it.

The repository

github.com/lerugray/generalstaff

AGPL-3.0-or-later, version 0.17.0. The orchestration, the gate, co-seats, Aides and the audit trail are all in it.

View on GitHub

Run one cycle

curl -fsSL https://raw.githubusercontent.com/lerugray/generalstaff/master/install.sh | bash
gs welcome

On Windows, run install.ps1 from the same place. The wizard asks about five questions and runs one verified cycle.

GeneralStaff Workbench is the desktop surface for the same engine: pick a project, pick which AI does the job, and direct the work in plain English.

06 · Early access

Join the list.

GeneralStaff is pre-launch. Leave an address and you will hear when seats open. No prices today. The list hears first.

A mock form for now. Nothing leaves this page.