Compound your coding agents. Bring any CLI coding agent at full power and activate Teams as parallel Rooms. Each Room is a separate filesystem workspace with replicas of its Team's agents and clones of its assigned repositories. Parallelize the work. Keep the thread. Share the setup.
Install with a Coding Agent · Platform support · Manual install · ▶ Quickstart
Paste this prompt into a Coding Agent you already trust. It inspects first, shows you an exact plan, and waits before changing your machine.
Install AgentsCommander safely using only the official `mblua/AgentsCommander` repository, its latest stable GitHub Release, and the canonical guide: https://github.com/mblua/AgentsCommander/blob/main/docs/install-with-agent.md
First resolve/report the current `main` snapshot's full commit SHA, then read the guide at that pinned commit. Treat the entire pinned guide, especially `Support gates`, `Use only pinned official evidence`, `Storage, backup, and upgrades`, and `Coding Agent contract`, as binding. Fail closed on every stop condition and on missing, ambiguous, or conflicting required evidence.
Before any download or mutation, perform/report every required read-only preflight and the complete plan: OS/version/architectures/support tier and Linux distribution; existing install/version/provenance; independently resolved target stable tag; storage, backup, upgrade, and uninstall-state requirements from the canonical owner https://github.com/mblua/AgentsCommander/issues/1118 (`main` is not published-binary evidence); mapped release/asset URLs and exact checksum record; destination, commands, created/overwritten files, privilege, PATH/system-wide effects, preservation, validation, and rollback. Apply the guide's gates exactly: Windows 10 1809+/11 x86_64/AMD64 is full; Linux x86_64/AMD64 is supported with medium test coverage and requires the limitations explained plus explicit confirmation; macOS is supported via npm only with very low test coverage, so warn that errors are expected, get explicit confirmation, and use no other macOS route; all other platforms stop. On Linux, identify the distro from read-only OS information using ID first, then recognized ID_LIKE family tokens: Debian/Ubuntu selects DEB, Fedora/RHEL selects RPM, other Linux selects AppImage. Conflicting family evidence or a missing exact mapped asset stops. Prefer least privilege. Await plan approval plus separate consent for elevation, system-wide/PATH changes, overwrites, and unsigned execution.
Only afterward follow the guide's exact same-release download, one-record full-SHA-256 verification, data preservation under #1118, approved execution, exact-path validation/reporting, and rollback rules. Stop an existing-install update or uninstall if canonical guidance is insufficient; do not invent storage or migration behavior. Warn that Windows artifacts and the macOS `.app` bundle may be unsigned, and that a matching release checksum detects corruption/asset mismatch, not publisher/repository-account compromise. Enforce all prohibitions: silent security bypass, automatic elevation, `curl | shell`, mirrors, source fallback, emulation/substitute assets, npm fallback outside the macOS route, and installing/authenticating a Coding Agent CLI for me.
Read the complete installation contract, platform matrix, trust boundaries, and rollback rules. Prefer a manual release download, or npm? See the other routes, including the npm-only macOS route, under the same platform policy.
See which Team is working on each task, even while several Teams work in parallel. Define agents and Teams in source-controlled files, then activate a Team as one or more Rooms. Each Room is a separate filesystem workspace with its own agent replicas, repository clones, messaging area, and build or test state. The UI keeps every Room's task, sessions, branches, and repository state visible.
- Pick the coding agent per role (Claude Code, Codex, Antigravity, Pi, or OpenCode) at full power. Each runs in its own real terminal with a full PTY, not a command runner. AgentsCommander only adds capability; it never sandboxes or nerfs your agent.
- Direct multiple Rooms from the Root Agent. The Agent Commander / Root Agent gives you one place to steer work across Teams. Ask it to talk to Room orchestrators, send work to different Teams, and keep initiatives aligned across parallel Rooms.
- Give each Room its own working copies. AC creates a separate workspace with repository clones, agent replicas, a messaging area, and build or test state. Each agent's context carries write-restriction instructions that ask it to stay inside its Room; that is documented policy the agents follow, not a sandbox the product enforces. Each session invokes the installed AgentsCommander executable with credentials whose token and root values are scoped to that Room. Run multiple Teams, or multiple Rooms for the same Team, without mixing one task's source changes or build and test state into another's.
- Return without rebuilding the map. A Room keeps its task, agent replicas, repository clones, and messages on disk. Come back later, reopen a Room's closed session, and AC asks Claude Code, Codex, Antigravity, or Pi to continue the prior conversation in that same workspace.
- Share configuration as source code. A Project's
.ac/tree holds canonical agent roles, skills, and plans; Team rosters and repo access; and Loop prompts and schedules in reviewable Markdown, JSON, and TOML. Commit the shared files so teammates can pull the same reviewed setup instead of recreating it. - Coordinate Teams through auditable files. Agents exchange Markdown messages in a
messaging/folder you can read withcat, inspect in an editor, search, and audit directly on disk. The whole organization fits inls. - Phone-ready updates with images. The Telegram bridge can stream session output and send photos or screenshots captured by agents, so remote status can include the actual screen or report.
- Local state, no telemetry. Shared team state stays in each project's
.ac/tree. Follow the canonical storage, backup, upgrade, and uninstall-state contract before moving or deleting application state.
You bring the coding agents. AgentsCommander coordinates them.
- Install and validate AgentsCommander with the reviewable Coding Agent plan. If you choose a secondary route, follow the canonical platform and trust rules. Start the validated executable from its approved location.
- Open a project: click
New Projectin the sidebar and point it at an empty folder. AC creates a Project AC Root (.ac/) there. - Create a Team: add an orchestrator and one worker agent, each with a role prompt. Teams and rooms walks through this.
- Launch the orchestrator: pick Claude Code, Codex, Antigravity, Pi, or OpenCode from the dropdown. Ask it to send the worker a hello message. The worker terminal receives a file notification and responds in real time.
Full walkthrough: docs/quickstart.md.
Most agent tools focus on in-process orchestration or one interactive session. AgentsCommander starts with the coding agents you already use (Claude Code, Codex, Antigravity, Pi, and OpenCode), runs them as real OS processes, and lets them coordinate through plain Markdown files that humans and file tools can read, search, and audit directly on disk. You see every step in a real terminal, and the coordination state stays visible on disk.
| Use case | Setup |
|---|---|
| Parallel feature development | Activate one Room per parallel task, even when reusing the same Team. Every Room gets its own repository clones and build or test state; direct the Room orchestrators from the Root Agent. |
| Code-review swarm | One agent ships a PR; two others review independently. You read both reviews in their own terminals before merging. |
| Autonomous refactor Team | A long-running orchestrator splits a multi-file refactor across worker agents and rebases their branches as they finish. |
| Long-running agent with phone alerts | Pair a session with a Telegram bot, kick off a build from your phone, and receive text updates plus screenshots or image artifacts. |
Full recipes: docs/use-cases.md.
| AgentsCommander | LangGraph | AutoGen / AG2 | CrewAI | Aider | Claude Code alone | |
|---|---|---|---|---|---|---|
| Operates real CLI coding agents | ✅ Claude Code, Codex, Antigravity, Pi, OpenCode | ❌ Python LLM calls | ❌ Python conversation | ❌ Python library | Partial (one agent) | ✅ (one agent) |
| Real PTY per agent | ✅ ConPTY / Unix PTY | ❌ | ❌ | ❌ | ✅ | ✅ |
| Filesystem-first messaging | ✅ Markdown in messaging/ |
❌ DB / Python state | ❌ Python objects | ❌ Python tasks | n/a | n/a |
| Standalone runtime | ✅ Rust / Tauri | ❌ Python library | ❌ Python library | ❌ Python library | ❌ Python app | ✅ standalone CLI |
| Multi-agent on the same repo | ✅ | Partial | Partial | Partial | ❌ | ❌ |
| Desktop UI | ✅ Tauri app | ❌ | ❌ | ❌ | TUI only | TUI only |
Full comparison with trade-offs and honest losses: docs/comparison.md.
These are not accidents.
- Start with files and CLIs. AgentsCommander keeps the core workflow in plain files and real terminal sessions.
- Configuration is source code. Canonical agent, Team, and Loop configuration lives in plain Markdown, JSON, and TOML under each Project's
.ac/tree. Commit those shared files and review them like code. Gitignored runtime and machine-local state (Rooms, Loop state, and project-local settings) can also live under.ac/; keep it distinct from the versioned configuration. - Files before databases. Shared configuration and inter-agent messages use plain files that are easy to inspect and debug. Databases can be introduced for performance-critical runtime paths without hiding how a Team is defined or how its agents coordinate.
- One agent = one directory. An agent is defined by a
CLAUDE.mdfile (or equivalent role-prompt file) inside its own directory. Multiple role prompts within the same directory or its subdirectories are forbidden. Coding agents assume the entire contents of their working directory are relevant context; if two role prompts coexisted, an agent could read another agent's role and leak context.
- Install with a Coding Agent: platform gates, approved assets, checksum verification, and rollback
- Quickstart: installed app to first running agent
- Concepts: agent, team, room, orchestrator, brief
- Teams and rooms: orchestrators, members, briefs, messaging
- Features: Coding Agent Profiles, Session auto-close, Config seed, Seed manifest, Container coding agents, Telegram bridge with image and screenshot sends, voice-to-text, portable instances
- Reference: full CLI,
settings.jsonschema, architecture, log filtering - Roadmap · Changelog · Docs style guide
| Platform | Native architecture | Status |
|---|---|---|
| Windows 10 1809+ / Windows 11 | x86_64 / AMD64 | Fully supported; primary development and release-validation platform. |
| Linux | x86_64 / AMD64 | Supported with medium test coverage; some features are untested or unsupported, so expect some turbulence. Confirm the limitation before installing. |
| macOS | Any | Supported via npm only, with very low test coverage; errors are expected. |
| Other combinations | Any | Unsupported; do not substitute assets or use emulation as an install fallback. |
An artifact on a release is not a support promise. The canonical installation guide defines the operational gates and lists verified Windows-only features.
AgentsCommander does not collect telemetry, analytics, or usage data. Optional features (Telegram Bridge, Voice-to-Text, and Jev classification for Co-managed rooms, which sends captured agent text only when the global coManagedEnabled switch, off by default while in development, is on, a room's flag is on and a Jev key is set) transmit data to external services only when you enable them; see PRIVACY.md. Windows code signing is planned through SignPath Foundation with free signing courtesy of SignPath.io, but current Windows release artifacts may be unsigned until epic #717 is complete. Verify the exact downloaded asset against the same release's SHASUMS256.txt; on Windows you can inspect signature status separately with:
Import-Module (Join-Path $PSHOME "Modules\Microsoft.PowerShell.Security\Microsoft.PowerShell.Security.psd1") -ErrorAction Stop
Get-AuthenticodeSignature -LiteralPath ".\Agents.Commander_<version>_x64-setup.exe"A release checksum detects corruption or a file that differs from the release record. It does not protect you if an attacker can replace both the asset and checksum through a compromised publisher or repository account. Full verification steps are in the installation guide; signing policy is in CODE_SIGNING_POLICY.md.
- Questions, ideas, show-and-tell: GitHub Discussions.
- Bug reports, feature requests: GitHub Issues.
- What's next:
ROADMAP.md.
Linux is supported with medium test coverage and macOS via npm with very low test coverage. Both tiers need the reports that turn gaps into fixes. If either platform misbehaves, file a reproducible GitHub issue with your OS version, native architecture, AgentsCommander version or exact release asset, exact steps, expected result, actual result, and relevant sanitized logs. Contributors can follow CONTRIBUTING.md; the installation guide has a copyable report template.
See CONTRIBUTING.md: branch naming, local build, log-filter setup, and the docs style guide.
AgentsCommander stands on the shoulders of:
- @msitarzewski/agency-agents: community library of agent role templates. AC can download an explicit, offline cache with
agency-templates update; normal startup and role browsing never hit the network. Big thanks to the maintainers for keeping it open.
Also: Tauri, SolidJS, xterm.js, portable-pty, axum, tokio; the toolchain layer this app would be impossible without.
Mariano is passionate about software development, AI, and blockchain. He approaches the world with deep curiosity, always amazed by life and the universe. Above all, he is a father, which remains the most wonderful part of his life.
Mariano Blua: GitHub · LinkedIn · 🇦🇷 MarianoBlua · 🇺🇸 MarianoBluaEN (English)