multi-agent-starter plugin
A generator that builds a file-based multi-agent system in a chosen folder: orchestrator, roles, approval gate and self-check.
- 106 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Workers launch external models and Codex runs commands in a `workspace-write` sandbox
- The generator overwrites system files in the target folder
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/multi-agent-starterDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Workers launch external models and Codex runs commands in a `workspace-write` sandbox
- The generator overwrites system files in the target folder
This is third-party code. Review the repository files before installing.
What it does
multi-agent-starter does not run agents itself but creates a scaffold in which they can run. The "configure a multi-agent system" flow asks for a flavor (`claude`, `codex` or `antigravity`) and a target folder, then copies ready templates there. The orchestrator differs per flavor and workers are called roles: `claude-main`, `codex-main`, `codex-critic`, `gemini`. The generated system includes an approval gate before every external model call, a `_shared/backends.json` registry mapping roles to models and connection modes (native, MCP, CLI, API), the `call_worker.sh` dispatcher, a protocol for resuming work after a pause and four topologies: pipeline, fan-out and fan-in, expert pool, producer and reviewer. State lives only in files, and `validate.py` checks the structure right away and reports PASS or FAIL. Generation is deterministic: templates are copied without model creativity.
Who it is for. Developers who want to tie Claude Code, Codex and Gemini into one team with explicit approvals and file-based memory.
Good fit when
- Work has to be split between an orchestrator and several worker models.
- Explicit approval is required before every call to an external model.
- A worker model or connection mode is swapped without touching system rules.
- A session breaks off and work must resume from records in files.
Not a fit when
- One agent without roles and approvals is enough.
- A ready cloud platform is needed: the project creates local files and has no background service.
- The environment is Windows without WSL: the Gemini worker is not verified there.
Example request
Set up a multi-agent system of the claude flavor in ~/agents/research, then create a task and run the validate.py check.Limitations
Developed and verified on macOS and Linux. Windows is experimental: installation, the Claude worker and the Codex worker work, while the Gemini worker needs a POSIX environment such as WSL and is not verified yet. Workers need Python 3, `jq`, `git` and the CLIs of the corresponding models (`codex`, `agy`), and Codex runs sandboxed through `codex exec`. The README and documentation are in Korean. The generator overwrites system files on update, so commit local edits to Git first. The repository was last updated on 23 October 2026.
How to disable. Delete the generated system folder or only its service files (`CLAUDE.md`, `_shared/`, `_templates/`) and disable the plugin in the `/plugins` list. Keep data from `tasks/` and `_local/` separately if you need it.
FAQ
Which orchestrator flavors exist?
Three: `claude` with Claude Code, `codex` with Codex and `antigravity` with an Antigravity session.
What is the approval gate?
A rule under which explicit user consent is requested before every call to a worker, meaning an external model.
Can a worker model be changed?
Yes. The `_shared/backends.json` registry links a role to a model and connection mode while system rules stay the same.
How to verify the generated result?
After installation `validate.py` runs automatically: it checks structural invariants and prints PASS or FAIL.
What happens with an existing folder?
The generator works in update mode: `tasks/` and `_local/` data are preserved while system files are overwritten with templates.
Discussion
No questions or reviews yet.
Related
PluginsA scaffold of personas, rules and skills in plain Markdown: Maestro hands work to different models and spares your main plan
multi-agents runs a team of coding agents on a task and measures whether it beat a single agent and what it cost.
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
Claude Code plugin with 60 subagents and 62 team presets, plus a browser designer for planning agent teams visually.
One agent launches others in the background: Claude Code, Codex, Copilot CLI, Grok Build and Pi, with a model chosen per stage.
A LangWatch command-line tool that creates the scaffold of a new AI agent project and hands it to a coding assistant
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| multi-agent-starter (this card) | no | 106 | |
| Claude Code Multi-Agent | no | 307 | a working folder with hooks and local Ollama that picks skills; here a generator of roles and an approval gate for three orchestrators |
| Harness Starter Kit | no | 114 | turns recurring agent mistakes into repository rules; here a scaffold for a team of several models |
README in short
The README describes the three flavors, four topologies, self-check and the tool list in detail, and Windows limitations are moved to `KNOWN_ISSUES.md`.