A software team of AI agents that works in your repo
The One's Software team takes a feature, a bug or a red test suite and works in your project folder, on your computer, with your own toolchain. A lead builds while checkers, started beside it, turn your request and the project's conventions into checks, and a developer takes tasks from a shared task board. The work comes back as a delivery whose tests, diffs, commands and spend come from the app's own records, and you accept it, send it back for rework or give it up.
What you can hand over
- “Fix the failing tests and say what caused them.”
- “Add CSV export to the reports page, with tests.”
- “Move the settings screen to the new form library and keep the build green.”
Say it to the main assistant in your own words. It opens a matter for the work, with its own board, and passes your words to the team as you wrote them.
How the team works
- One owner, independent checks. The lead builds. The checkers get your request, the conventions and the interfaces, not the lead's code, so they do not inherit its mistakes. The developer works through tasks on a TaskBoard, where a reviewer checks each one.
- It asks you, not itself. When the team needs a decision, the question waits for you; the main assistant never invents your answer. You can also talk to the team at any time.
- It keeps going. The work lasts as long as it takes, across days and app restarts. A member that hits an error retries by itself.
- Any model, per role. Each role can run on a different model from any vendor you have connected. Ask for a stronger model for one piece of work, or set each role's model in a Space of your own.
What comes back
A delivery checked against what the matter asked for, with an evidence pack built from the app's records, “not from the Agent”: test output, diffs, the commands that ran, any outside actions, and what it spent. Accept, Rework or Give up.
What gets recorded
The team works in your folder with your user account's rights. Every tool call is recorded, shell commands included, and file changes can be undone: one call, a run, or all of a matter's work. Pushes and other actions that leave your computer are marked as ones that cannot be taken back. Keep using Git as usual: dependency, cache and gitignored folders are not restored.
Other ways to start it
- Linux server:
the-one-server run "Fix the failing tests and say what caused them" --workdir .runs the Software team headless and writesdelivery.md,run.jsonandtrajectory.json. - From Claude Code or Codex: connect The One's MCP server and hand a job to the team without leaving your terminal. See The One vs Claude Code and The One vs Codex.
Where it stops
- It is not an editor plugin and has no inline completions.
- Agents stop when your computer is off, unless they run on your Linux server.
Questions
- Which models can the team use?
- Any you connect: the official service, a ChatGPT sign-in, OpenAI, Anthropic or Gemini keys, or a compatible endpoint, local ones included. Each role can use a different one.
- Can it push to my remote?
- It works with your rights, so yes if the work calls for it. A push is recorded and marked as an action that cannot be taken back; say in your request if it should not push.
- What does it cost?
- Nothing, with your own keys or ChatGPT plan: those calls are billed by the provider. A membership ($20, $60 or $200 a month) includes the official model service, so you need no keys at all.
In the docs