Best Open-Source Agent Orchestrators for Coding (2026)

Cloudroom is the best open-source agent orchestrator in 2026, because it runs Claude Code, Codex and more agents in parallel on your own subscription, and its app is Apache-2.0. The other strong picks are Emdash for per-task environments on your own infra, T3 Code for reaching your agents from any device, Claude Squad for a minimal tmux app, Agent Orchestrator for hands-off PRs, Paseo for building on top of it, and Herdr for a keyboard-first terminal multiplexer.

An agent orchestrator is an app that runs several coding agents, such as Claude Code and Codex, in parallel, each in its own worktree or sandbox.

TL;DR

• Best overall: Cloudroom, open source with hosted cloud agents.
• Best for per-task environments on your own infra: Emdash, free and Apache-2.0.
• Best for reaching your agents from any device: T3 Code, with desktop, web, iOS and Android apps.
• Best minimal terminal app on top of tmux: Claude Squad, free and AGPL-3.0.
• Best for hands-off PRs: Agent Orchestrator, which sends failed CI checks and review comments back to the agent that wrote the code.
• Best for building on top of it: Paseo, free, with plugins, a TypeScript SDK, MCP and the CLI.
• Best keyboard-first terminal multiplexer: Herdr, free and Apache-2.0.

How I picked them

Every tool here carries an OSI-approved license, and I checked each license on GitHub on October 16, 2026. Each one runs several existing coding agents in parallel, each in its own worktree or sandbox. That removes conductor.build, which is proprietary, and Superset, which uses the Elastic License 2.0.

I have used T3 Code, Herdr and cmux. The other sections come from each tool's docs and repository.

Open-source agent orchestrators compared

Last checked October 16, 2026.

1. Cloudroom: best overall

Best for: developers who want an open-source app and cloud agents on their own Claude or ChatGPT subscription.

Cloudroom is a desktop app for coding agents on macOS and Linux. Local threads run 9 agents, Cursor included. Cloud threads run Claude Code, Codex, Pi Agent, OpenCode, fx and Hermes, each in its own Linux sandbox and Git branch, on your own account with no token billing. The app and core are open source under Apache 2.0, and you can run the core on your own Linux server.

It ranks first because it is the only open-source tool here that hosts Claude Code and Codex in the cloud on your own Claude or ChatGPT subscription. cmux Cloud VMs run on a BUSL-1.1 server, so that part is not open source. OpenHands Cloud bills models through your own API key or its providers.

room-cli thread spawn --project PROJECT --machine cloud --provider claude-code \
  --model claude-sonnet-5-5 --request-id auditlog-01 \
  --prompt "Add cursor pagination to the audit log endpoint and open one PR."
room-cli thread tell THREAD_ID "Also update the API docs with the new cursor parameter."

The free plan gives you 100 cloud agents and 25 hours of cloud compute every 30 days, plus unlimited local agents. Pro is $50 a month. The honest con is that several people cannot work inside the same thread today, and there is no Windows app.

2. Emdash: best for per-task environments on your own infra

Best for: teams that want their own scripts to build a fresh environment for each task.

Emdash is an Apache-2.0 desktop app for macOS, Windows and Linux. It pulls issues from Linear, Jira, GitHub, Notion and Asana, and gives each task a Git worktree, an in-app browser, a diff, PR and CI review, scheduled automations and remote SSH. Its .emdashrc setup scripts provision each task's environment on infrastructure you run. The site claims over 1M downloads. Emdash is free.

The genuine pro is that per-task setup. Pick something else if you want the cloud hosted for you.

3. T3 Code: best for reaching your agents from any device

Best for: developers who want to check and steer their agents from a desktop, a browser or a phone.

T3 Code is a free, MIT-licensed control plane for coding agents. It ships desktop apps for macOS, Windows and Linux, a web app, and iOS and Android apps. Providers are Codex, Claude, Cursor, Grok, OpenCode and Antigravity. Its Device panel shows a live iOS Simulator or Android Emulator next to a thread, and agents control it through the device_* tools and agent-device command.

I tested T3 Code. Onboarding imported my Claude Code and Codex chats from the last 30 days. I signed up for cloud agents and only got a T3 Connect account, a tunnel. Agents run on the machine that hosts your T3 Code server. The genuine pro is the reach: desktop apps for all 3 systems, a web app, and iOS and Android apps. The Device panel is a bonus for mobile developers. Pick something else if you need hosted cloud.

4. Claude Squad: best minimal terminal app on top of tmux

Best for: terminal users who want the smallest possible app for several agents at once.

Claude Squad is an AGPL-3.0 terminal app called cs. It manages Claude Code, Codex, Gemini, Aider and other local agents, each in its own Git workspace. It needs tmux and the GitHub CLI, and you install it with Homebrew or a shell script. The last release was v1.0.20 on August 20, 2026, and it is free.

It stays small because tmux does the heavy lifting. Pick something else if you want a GUI, a diff review panel or hosted cloud.

5. Agent Orchestrator: best for hands-off PRs

Best for: teams whose agents should close the loop from a failed check back to a green PR.

Agent Orchestrator is an Apache-2.0 desktop app for macOS (Apple silicon and Intel), Windows and Linux. The repository moved from ComposioHQ/agent-orchestrator to OrchestratorInc/agent-orchestrator, and the product now lives at orchestrator.inc. Every project gets a main agent that plans the work and spawns workers into their own worktrees. The app runs its own daemon, so no CLI is needed, and it supports 25 harnesses including Claude Code, Codex, Cursor and Goose.

Its genuine pro is the feedback loop. It watches every PR it opens, and failed checks and review comments go back to the session that owns the branch until the PR is approved. A mobile companion pairs over LAN or Tailscale, and code stays on your machine. The site lists no price. Pick something else if you want hosted cloud.

6. Paseo: best for building on top of it

Best for: developers who want to extend and automate their orchestrator, not only use it.

Paseo is a free, Apache-2.0 control plane for coding agents. It runs Claude Code, Codex, OpenCode, Pi, OMP, Cursor, Muse Code, Antigravity and 33 more. There is a desktop app with no server required, a native mobile client, remote machines, and triggers from GitHub, Slack and Discord. You automate it with MCP, the CLI or a TypeScript SDK, and plugins add panels, slash commands, themes and new agent providers.

The genuine pro is how much surface it exposes. One person maintains Paseo full time, with no company and no investors. It has no telemetry, tracking or forced login. Pick something else if you want a company behind the tool or hosted cloud.

7. Herdr: best keyboard-first terminal multiplexer

Best for: terminal-fluent developers who want a multiplexer, driven by the keyboard, whose terminals keep running after they disconnect.

Herdr is a terminal multiplexer whose background server owns the terminals. It ships one binary for macOS, Linux and Windows, detects 22 agents, and is open source under Apache-2.0. Add other machines over SSH with herdr machine add, and they keep working when you disconnect. Version 0.9.5 landed on October 14, 2026. Herdr is free. Herdr Cloud is still "coming soon".

I have used Herdr a few times. It gets efficient once you learn the shortcuts, and rough without them. The docs call Herdr mouse-first and say you do not need shortcuts to start, which did not match my experience. The genuine pro is the one binary and the SSH machines. Pick something else for hosted cloud.

Other open-source agent orchestrators worth knowing

Nimbalyst. A free, MIT-licensed visual workspace with markdown, diagrams, Excalidraw and a session kanban beside parallel sessions in worktrees. It runs Claude Code, Codex, OpenCode and Copilot on macOS, Windows and Linux, plus an iOS app. Teams cost $20 per user per month and are free during beta.

Bernstein. An Apache-2.0 governance layer for AI agents. A deterministic Python scheduler with no model in the coordination loop plans work, spawns agents into worktrees, and a "janitor" verifies tests, types and lint so only verified diffs merge. It supports 53 CLI adapters, and the README says it is beta.

cmux. A native macOS terminal for coding agents, built on the Ghostty engine, with a TUI on macOS, Linux and Windows. The app is free and GPL, and Cloud VMs need Pro at $50 a month. The cmux server is BUSL-1.1, so the cloud part is not open source.

Vibe Kanban. A kanban board for 10 or more agents, including Claude Code, Codex, Gemini CLI, Copilot, Amp and more. It is Apache-2.0, but the README says "Vibe Kanban is sunsetting", and the last release was April 24, 2026. Do not set it up for new work.

OpenHands. An MIT-licensed agent platform that can drive Claude Code, Codex, Gemini CLI, its own OpenHands agent or your own stack. OpenHands Cloud is a hosted SaaS with a free Individual tier, where you bring your own key or use its providers at-cost. It also has automations from GitHub, Slack, Jira, Linear and schedules.

Which one fits you

• You want cloud agents on your own subscription: Cloudroom. On Windows, use T3 Code or Emdash.
• Your own scripts should build each task's environment: Emdash. For hosted cloud, use Cloudroom.
• You want your agents on every device you own: T3 Code. For hosted cloud, use Cloudroom.
• You want the smallest terminal app: Claude Squad. For a keyboard-first multiplexer, Herdr does more.
• Your agents should fix their own failed CI: Agent Orchestrator. For per-task environments, use Emdash.
• You want plugins and an SDK: Paseo. Claude Squad stays simpler.
• You live in the terminal and run your own server: Herdr. For managed cloud, use Cloudroom.
• You want diagrams and kanban beside sessions: Nimbalyst.
• You want policy and audit records: Bernstein.
• You want a native macOS terminal with cloud VMs: cmux.
• You already run Vibe Kanban: keep it, but do not start there.
• You want an MIT agent platform with its own agent and a hosted SaaS: OpenHands.

The short version

Cloudroom is my pick for best overall, because it is open source and runs cloud agents on your own subscriptions. The other six each win a specific case. I work at Cloudroom, so weigh that as you read.

You can create a free Cloudroom account, read best apps for running multiple Claude Code and Codex agents, the Conductor alternatives, or cloud coding agents compared.