Free and open source coordination for real CLIsTurn coding agents into repeatable teams.
AgentsCommander coordinates the coding-agent CLIs and sessions you configure in real terminal sessions. Define a coordinator and specialist workers once, then launch isolated workgroups for each task with local markdown handoffs you can inspect incat orgit diff.
AgentsCommander is free and open source under the MIT license. Bring your own coding-agent CLIs, accounts, and model access.
First-class profiles cover Claude Code, Codex, and Gemini CLI. OpenCode works through the configured coding-agent path, and Cursor CLI, Hermes, or other terminal agents can be added as custom commands when they run from the shell in your environment. AgentsCommander spawns the local CLI processes you provide. It does not bundle third-party tools, models, subscriptions, API keys, accounts, or provider access.
Example setup: user-provided CLIs and models. Accounts, API keys, model availability, and pricing are managed with the respective vendors.
- coordinatorCodex profile, GPT 5.5 xhigh
- dev-rustOpenCode configured CLI, GLM 5.2
- dev-uiClaude Code profile, Opus 4.8 max
- reviewerCustom CLI, Hermes + DeepSeek V4 Pro
Share the team setup, not a prompt snippet.
AgentsCommander fits teams that need the same multi-agent setup again and again. Define the coordinator, worker roles, repo access, and launch settings once, then activate matching workgroups for each new task.
- 01
Define once
Set the coordinator, worker roles, repo access, and launch settings a single time.
- 02
Activate a workgroup
Spin up a matching workgroup the moment a new task needs that same team.
- 03
Repeat
Reuse the same setup across features, reviews, refactors, and test loops.
Concrete capabilities, not abstractions.
Every step of multi-agent work stays visible: real terminals, explicit handoffs, and plain files you can read with the tools you already have.
Define the team once
Create a coordinator plus specialist workers once. Reuse the same Team for features, reviews, refactors, documentation passes, or test/fix loops.
Launch a clean workgroup for each task
Activate a Team as a Workgroup when work starts. Each Workgroup includes its own
TASK.md,messaging/folder, agent replicas, and separate repo/work state.Run real coding agents in real terminals
AgentsCommander coordinates the CLI tools developers already use instead of hiding them behind a framework abstraction. Each session runs in a real PTY, so prompts, failures, logs, and progress stay visible.
Delegate through a coordinator
Give the coordinator the goal. It can split the work, send markdown assignments to worker agents, collect replies, and move the task through explicit handoffs.
Inspect every handoff
Every inter-agent message is a timestamped markdown file on disk. Read it, diff it, archive it, or debug it with normal developer tools.
Clear context with a handoff
For long-running agent work, an agent can write an explicit handoff, clear its own context, and resume from the important task state captured in that file.
Repeat launches with profiles and config
Use Coding Agent Profiles and config seed to help keep role-specific launch parameters, environment settings, and replica setup consistent across repeated workgroups.
Watch parallel work in one place
Track coordinator and worker sessions in the desktop UI, see running/waiting/exited states, and keep parallel agent work visible while it happens.
Add optional remote control for long runs
For long-running work, Telegram and voice features can help you monitor or send prompts while away from the keyboard. Treat this as optional: Telegram and Gemini voice transcription use external services when enabled.
Put a repeatable team on real work.
The same setup drives many jobs. You stay in the loop at every handoff.
Parallel feature development
Two coding agents work the same repo, each owning a different module. The coordinator routes the work and brings the results together for you to review.
Code-review swarm
One agent opens a pull request while two others review it independently. You read each review in its own terminal before you merge.
Refactor across workers
A coordinator splits a multi-file refactor across worker agents and tracks each branch as the workers finish, with every change visible on disk.
Test and fix loops
Run a build or test pass, route the failures to a worker to fix, then repeat the loop with each step visible in a real terminal.
Documentation pass
Point a worker at a module to draft or update its docs while the coordinator keeps the change set organized.
Long run with phone alerts
Pair a session with the optional Telegram bridge, start a long task, and receive updates plus screenshots while away from the keyboard.
Every handoff is a file on disk.
Agents talk to each other through timestamped markdown in a messaging/ folder. It is an inspectable local markdown trail, so you can read it, diff it, or archive it with the tools you already use. No hidden state, no database to query.
- cat a message to see exactly what one agent told another.
- git diff the folder to watch the coordination evolve.
- grep the trail when you need to debug a run.
Your project state stays on your machine.
AgentsCommander keeps project state and inter-agent messages on disk as plain JSON, TOML, and markdown. It does not collect telemetry, analytics, or usage data.
Optional features are opt-in: the Telegram bridge and Gemini voice transcription send data to those external services only when you turn them on.
- WindowsPrimary platform, where most development happens.
- LinuxSupported, testing is underway.
- macOSNot yet validated.
Windows releases are digitally signed through SignPath.
Create your first team.
Install the CLI, open a project, add a coordinator and a worker, then launch them in real terminals. The quickstart walks the whole path.
$ npm install -g @mblua/agentscommander
$ agentscommanderThe published npm package is @mblua/agentscommander. The installed command is agentscommander.
Prefer a desktop install? Get the signed Windows installer, Linux AppImage, or macOS build from GitHub Releases. You bring and authenticate your own coding agent CLIs.