Larceny is a Claude Code plugin that gives you a small software team made of AI agents. You talk to one coordinator. He breaks your request into small tickets, hands them to coding agents who work in parallel, and makes sure an adversarial reviewer checks every pull request before it merges. Every change has a ticket, a branch, a pull request and a review thread on GitHub, so you can trace it back later.

The default crew is named after characters from the TV series Prison Break, with a reviewer from Brooklyn Nine-Nine and an advisor from Star Wars. You can rename any of them.

Every persona is an AI agent, and agents make mistakes. They can state something wrong with confidence, miss a real problem in review, or act on bad information. Review their work, check claims against the code or GitHub rather than an agent's report, and check before anything you can't undo: a merge, a release, a delete. A team runs several model sessions at once, so it uses your plan's limits faster than one agent. See docs/cost-and-safety.md.

- Claude Code.

- A git repository with at least one commit and a GitHub remote. Each coder works in its own worktree, so the repo needs a main branch to branch from.

- The GitHub CLI, gh, signed in to an account with access to that repo.

- A project board. GitHub Projects and Linear (through its MCP server) are both tested.

- Optional: a GitHub account per persona, so each one shows up under its own name. See Persona accounts.

- Optional: Discord, to talk to the coordinator from your phone.

Inside Claude Code:

/plugin marketplace add nestedmind/larceny

/plugin install larceny@larceny

Or ask Claude:

Install the Larceny plugin from the nestedmind/larceny marketplace using the claude plugin CLI, then tell me if I need to restart.

Inside Claude Code, in your project:

/larceny:onboard

Or ask Claude:

Onboard this project with Larceny.

Onboarding checks your gh login, repo and access, asks who you are and how you want reports, agrees the test, lint and build commands with you, and lets you pick names and models for the crew. It saves your answers in .larceny/config.md and adds .larceny/ to .gitignore. A board, a branch ruleset and persona accounts are offered last, and you can skip each one. Running it again shows your answers and asks before changing any.

/larceny:wake-up

Or ask Claude:

Wake up Larceny's coordinator for this project.

The session now acts as Scofield, or as your renamed coordinator. If the project hasn't been onboarded, he runs onboarding first. He asks before starting any agent.

From a terminal, start a session as the coordinator:

claude --agent larceny:coordinator

To make the coordinator the default for a project, add {"agent": "larceny:coordinator"} to .claude/settings.json. The plugin never sets this for you.

The models are a recommendation. Onboarding lets you keep them, run everything on your harness default, or change any persona.

- You tell the coordinator what you want.

- He turns it into tickets on the board and works out which ones can run at the same time.

- Each coder takes one ticket, builds it on its own branch with tests, and opens a pull request.

- Amy reviews the pull request against its ticket. The coder fixes what she raises.

- If a review reaches a third round, the coordinator steps in. At the fifth round, it is escalated to you.

- Once Amy passes it, the work merges and the coordinator reports back.

Add a project name after a spawn command, such as /larceny:spawn-advisor my-app, to give the persona one project to look at. Once a persona is running, ask the main session to relay: "Ask Amy: review PR 12." You can also start a session straight into a role with claude --agent larceny:reviewer, larceny:advisor or larceny:teacher. Details are in docs/agent-lifecycle.md.

- Rename the crew. Onboarding can write your own names for any persona without changing the plugin's files. See docs/renaming.md.

- Use the same crew everywhere. Save your crew once and every new project on this machine can use it. A project's own config always wins. See docs/crew-resolution.md.

- Change models. Set per persona in the models:line of.larceny/config.md. See docs/agent-lifecycle.md.

One ordinary GitHub login is enough to run the team. Giving each persona its own account is optional. Without one, every commit and review shows up under your name, and because GitHub doesn't let an account approve its own pull request, Amy gives her verdict as a comment instead of an approval. A person has to create each account, since GitHub requires email verification and a captcha. See docs/identity-wiring.md.

Inside Claude Code:

/plugin uninstall larceny@larceny

Or ask Claude:

Uninstall Larceny. First list every file and setting it left in this project, using .larceny/config.md to find them, and wait for my OK before deleting anything.

Uninstalling removes the plugin but not the files its work left in your projects: renamed persona files, .larceny/, worktrees, tokens and settings. The prompt above finds them for you. To do it by hand, follow docs/uninstall.md.

Codex can read the skills in skills/ through .codex-plugin/plugin.json. The agents and commands are Claude Code features, so Codex users run the personas by hand. None of this has been tested. If you try it, tell us what worked on the issues page.

- docs/limitations.md: what is untested or doesn't work yet, including why a GitHub comment doesn't wake a local session.

- docs/smoke-test.md: what has been tested, and the checklist for a clean-machine install.

- docs/cost-and-safety.md: token cost, GitHub tokens and what agents can run.

- CONTRIBUTING.md: how to pick a ticket, open a pull request and add a skill or persona.

MIT. See LICENSE. Adapted skills are credited in THIRD_PARTY.md.