Short answer
Claude Code is the agent; Orchestrator.inc (formerly Agent Orchestrator) is a supervisor that runs Claude Code, and other agents, on top. If Claude is your only agent and two or three parallel sessions is your normal load, Claude Code's built-in worktrees, agent view and agent teams are enough, with nothing extra to install. Orchestrator earns its place when sessions outnumber your attention, when CI and review feedback should reach the right session by itself, or when you want a second vendor's agent reviewing Claude's work. Choose the built-ins if you run a few Claude sessions; choose Orchestrator if you run many sessions or many agents.
At a glance
| Orchestrator.inc | Claude Code built-ins | |
|---|---|---|
| Best for | Supervising many sessions across agents, with PR, CI and review state routed to the owner | Parallel Claude sessions with the least setup |
| Platforms | macOS, Windows, Linux, plus iOS and Android companion | CLI on macOS, Linux and Windows; Desktop Code tab on macOS and Windows, Linux in beta |
| Coding agents | 38 harnesses, including Claude Code, Codex, Cursor, OpenCode, Gemini CLI and Aider | Claude |
| Isolation | Git worktree and branch per session under ~/.ao; multi-repo workspaces | --worktree creates a worktree under .claude/worktrees/; agent teams share one checkout |
| Overview | Kanban board placed by session, PR, CI and review state | Agent view: a terminal list grouped by status; Desktop: a sidebar of sessions |
| PR and CI | Failing checks, review comments and merge conflicts sent to the owning session | Agent view shows a PR label; Desktop has a CI status bar with an auto-fix option |
| Agent interface | Structured Chat or the agent's native terminal UI | Terminal, or Desktop chat with a diff viewer |
| Planner agent | Project orchestrator that briefs and spawns workers | Subagents and experimental agent teams inside one session |
What Claude Code ships today
As of October 2026, Anthropic documents several ways to run Claude in parallel.
claude --worktree <name>creates.claude/worktrees/<name>/on its own branch, and a.worktreeincludefile copies gitignored files such as.env.- Agent teams are experimental and off by default. Teammates share the main session's checkout, and split panes need tmux or iTerm2.
- Agent view (
claude agents, a research preview) lists sessions in the terminal, grouped by status, with a PR label on each. - The Claude Desktop Code tab runs sessions in a sidebar with a worktree toggle, a diff viewer, PR creation and a CI status bar.
The problem in practice
You start the morning with three Claude sessions in three worktrees. By lunch you have seven, two of them waiting on CI, one with a review comment from a teammate, and one whose branch now conflicts with main. You also want Codex to give a second opinion on the biggest PR. Each fact is visible somewhere. The work is noticing it, finding the right session, and pasting the context in.
How each tool handles it
Claude Code on its own. Run claude --worktree fix-flaky-test, ask Claude to install dependencies in the fresh checkout, describe the failure, and let it open a PR. Watch it in claude agents or the Desktop sidebar. When CI goes red, the Desktop status bar can ask Claude to fix it; from the terminal you paste the failure back yourself. If the branch conflicts, you tell the session to rebase. A Codex review happens in another tool.
With Orchestrator.inc. Click New session on the project, pick Claude Code, and choose Chat or Terminal, or use ao spawn from the CLI. The worktree and branch are created under ~/.ao, and the project's postCreate and symlinks settings install dependencies and share .env. The PR is picked up once pushed. Failing checks, review comments and merge conflicts arrive in that session as messages. A Codex reviewer can run on the same PR and its findings go to the Claude worker. You press merge when the card reaches Ready to merge; merging is never automatic.
Where Claude Code's built-ins are a good fit
- Claude is your only agent and you rarely run more than a few sessions.
- You want everything inside your existing Claude subscription with no extra app.
- One problem splits into parts that should share one checkout, which suits agent teams.
Where Orchestrator fits better
- Mixed agents on one board. Run Claude Code next to Codex, Cursor, OpenCode or Gemini CLI. Claude Code and Codex sessions can switch between Chat and the native terminal UI and keep their worktree.
- A board that keeps itself current. Lanes (Working, Needs you, In review, Ready to merge) come from session, PR, CI and review facts, not drag and drop.
- Feedback that reaches the session on its own. GitHub and GitLab are polled with ETags, with no webhooks or GitHub App, and repeated messages are de-duplicated.
- Reviewers from another vendor. 22 of the 38 harnesses can act as reviewers on a worker's PR.
- Your phone. The iOS and Android app shows Chat, the terminal and diffs, with push when a session needs you or a PR is ready to merge.
- Windows users. The desktop app uses a native ConPTY terminal host, so no tmux is needed.
- Your own servers. Install Orchestrator's host on a server or cloud VM, then start and supervise sessions there from your desktop, with several hosts in one app. Machines linked to the same Orchestrator account find each other; otherwise they pair over an HTTPS tunnel or the LAN.
Pricing
Running Orchestrator on your own machines is free and open source (Apache-2.0), with no account needed; you bring your own agent subscriptions or API keys. Optional Orchestrator Cloud plans, for agents that keep working while your laptop is closed, are Starter at $20/month and Pro at $40/month, with Enterprise pricing on request. Orchestrator does not proxy model calls, so Claude usage still comes from your own Claude plan or API key. Claude Code's parallel features are included in paid Claude plans as of October 2026.
Who should pick which
Stay with Claude Code's built-ins if
- Claude is your only agent and two or three sessions is typical
- you want hosted Claude sessions from Anthropic
- you want several Claude agents to share one checkout on one problem
Add Orchestrator if
- you run Claude Code next to other agents and want one board
- you have more open PRs than you can watch
- you want a reviewer agent from another vendor on each PR
- you want to approve and reply from your phone while the desktop keeps working
- you want Claude Code sessions on your own server or VM, driven from your desktop
Try it
Follow the parallel issues guide to spawn two Claude Code workers on two issues and watch both PRs move across the board. For a vendor-neutral walkthrough, see how to run multiple Claude Code agents in parallel.
Frequently Asked Questions
Does Orchestrator.inc replace claude --worktree?
You do not need the flag inside Orchestrator.inc. It creates a worktree and branch per session under ~/.ao and launches Claude Code inside it. You can still use claude --worktree on its own for a quick one-off session.
Can Orchestrator.inc run Claude Code agent teams?
Orchestrator.inc launches a normal Claude Code session in each worktree, so subagents and the experimental agent teams setting still work inside it. Orchestrator's own planner is separate: the project orchestrator agent spawns independent workers, each in its own worktree.
Is Claude Code's agent view a Kanban board?
It is a status-grouped list in the terminal with a PR label per session, as of October 2026. Orchestrator.inc's board places each session by its agent state plus PR, CI and review facts, and shows failing check names, links and the end of each failed job's log for each card.
Do I need the Claude GitHub App with Orchestrator.inc?
No. Orchestrator.inc polls GitHub and GitLab directly and sends failing checks, review comments and merge conflicts to the owning session without webhooks or a GitHub App.
Does Orchestrator.inc work with Claude Code on Windows and Linux?
Yes. The desktop app ships for macOS, Windows x64 and Linux, and Windows uses a native ConPTY terminal host, so no tmux or WSL is required. Download it from orchestrator.inc and point a new session at Claude Code to try it.