multi-agents CLI
multi-agents runs a team of coding agents on a task and measures whether it beat a single agent and what it cost.
- 158 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Runs several coding agents that edit files and run commands in worktrees.
- The `--permissions bypass` mode disables sandbox checks and token spending multiplies with the number of agents.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/multi-agentsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Runs several coding agents that edit files and run commands in worktrees.
- The `--permissions bypass` mode disables sandbox checks and token spending multiplies with the number of agents.
This is third-party code. Review the repository files before installing.
What it does
The `multi-agents run` command takes a repository, a team spec and a task. Worker agents run in separate git worktrees: Claude Code, Codex or a built-in agent on any model through the OpenAI Agents SDK. The result is one patch and an itemised record of every call: tokens, cost, seconds. Your working tree stays untouched unless you pass `--apply`. Team specs: `solo` (the baseline agent), `review` (a coder and a reviewer scoring the patch), `best` (parallel workers and a judge), `plan` (a planner splits the task into subtasks). A benchmark compares a team with a single agent on fresh tasks. The repository also works as a plugin for Claude Code and Codex through the `/multi-agents:team` command.
Who it is for. Developers deciding whether a team of agents pays off on their tasks who want to see the price of every run.
Good fit when
- A single agent and a coder plus reviewer pair need comparing on your repository.
- Several workers fix a bug in parallel and a judge picks the best patch.
- An issue or pull request should produce a ready patch with a cost report.
- A per-run spending limit is set with `--max-cost`.
Not a fit when
- One agent is enough for a small change.
- The run would go on a working machine with `--permissions bypass`.
- Token spending cannot be limited at all.
Example request
Run the best team with two workers on this bug, show the patch and the run cost and do not apply it without my word.Limitations
Python 3.10 or newer, `uv`, git and at least one logged-in CLI, `claude` or `codex`, are required. The `--permissions bypass` mode turns off sandbox checks and is meant for disposable machines and containers. The built-in `sdk` agent has no permission system. The `--max-cost` flag stops new runs but is not a hard ceiling: runs already in flight finish and the total can overshoot. An unpriced run blocks further runs unless `--allow-unpriced` is passed.
How to disable. Remove the plugin with `claude plugin uninstall multi-agents@multi-agents` or the package with `uv tool uninstall multi-agent`. Delete worktrees and run records by hand.
FAQ
How do I install the plugin in Claude Code?
`claude plugin marketplace add BichengWang/multi-agents`, then `claude plugin install multi-agents@multi-agents`. In a session `/multi-agents:team` works.
Are API keys needed?
Not for the `claude` and `codex` modes: they reuse your existing CLI login. The built-in `sdk` agent needs access to the chosen model.
Does a run change my working tree?
No. Each worker uses its own git worktree and the patch is applied only with `--apply`.
How do I specify a team?
With a string like `best:workers=claude@haiku+codex@gpt-6-luna`; the options are `solo`, `review`, `best` and `plan`.
What does the --max-cost flag do?
It stops new agent runs once the given dollar amount is spent. It is not a hard ceiling: runs already in flight finish.
Discussion
No questions or reviews yet.
Related
CLI toolsA memory and coordination layer for coding agents: specs and issues live in git while agents work in parallel
Ten AI agents in tmux panes: a manager hands tasks to seven workers, progress shows on a dashboard, with Claude Code, Codex, Cursor and OpenCode supported
Multi-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
A tmux-based TUI that runs Claude Code, Codex and other agents in one list: status, quick prompts, worktrees and diff review
An open coding agent for the terminal, browser and desktop that works with models from OpenAI and Claude to Ollama
One agent launches others in the background: Claude Code, Codex, Copilot CLI, Grok Build and Pi, with a model chosen per stage.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| multi-agents (this card) | no | 158 | |
| Claude Squad | no | 8,738 | manages several agent sessions in a terminal; multi-agents runs a team on one task and compares it with a single agent on cost |
| Claude Code Multi-Agent | no | 305 | a ready multi-agent configuration; here team specs, a judge and per-run cost accounting |
README in short
The project compares teams of coding agents with a single agent, tracks cost, and works from the console and as a plugin for Claude Code and Codex.