Orca

Orca is a desktop application with a local runtime. Each repository has a base ref; each task creates a git worktree on its own branch, and the agent terminals, editor tabs and browser tabs of that task are scoped to it. Orca starts the chosen agent's own CLI in that directory and forwards the user's subscription credentials, so it does not proxy model calls itself.

Coordination between agents is manual by default: the person opens worktrees, launches agents, reads diffs and decides what to merge. An optional orchestration feature, which the docs label experimental, adds a Run (a namespace and inbox), Tasks with dependencies, Dispatches (one attempt of a task on a terminal) and supervised workers that report back with worker_done messages. The coordinator can block a task on a decision gate.

State lives in the local Orca runtime and in the git worktrees. The runtime can also sit on another machine (SSH target, Remote Orca Server or cloud VM the user owns), with the desktop or phone acting as a client. The README says the relay that pairs the mobile app with a desktop host is in the same repository under cloud/.

The human is in the loop by default: agents run in visible terminals, and diffs are reviewed in the app. New launches pre-fill each agent's permission-bypass flag (Yolo); Settings has an Agent Permissions switch to Manual, which restores the agent's own prompts. Pushing is explicit and Orca does not silently force-push. Notifications signal when an agent finishes or needs attention. In the experimental orchestration layer, decision gates block a task until the coordinator resolves them, and worker-stop ends a worker.

brew install --cask stablyai/orca/orca

# Desktop app first (README also lists .dmg, .exe, AppImage and AUR builds)
brew install --cask stablyai/orca/orca

# Orchestration is experimental: enable it under Settings > Experimental,
# then check that the CLI can reach the running app
orca status --json

orca orchestration run-create --objective "Split checkout QA and summarize blockers" --json
orca orchestration task-create --spec "Audit billing settings for mobile layout" --task-title "Billing audit" --json
orca orchestration worker-start --task <taskId> --worktree new-child --name billing-audit --agent codex --setup run --json
orca orchestration check --wait --types worker_done,escalation,question --timeout-ms 900000 --json

• The orca CLI talks to the running desktop app; orca status --json must succeed before orchestration commands work.
• Orchestration must be enabled under Settings > Experimental. The older orca orchestration run and run-stop commands are retired and do nothing.
• New agent launches default to each CLI's permission-bypass flag (for example --dangerously-skip-permissions for Claude). Switch Settings > Agents > Agent Permissions to Manual to keep prompts.
• On Linux the CLI binary is named orca-ide (for example orca-ide serve --pairing-address <host>).
• Each agent CLI must be installed and logged in separately; Orca launches it and forwards your credentials.

Partly. The docs describe registering MCP servers in Orca's settings so their tools show up in agent CLIs that support MCP. They do not describe Orca as an MCP server, and the record marks MCP as partial.

Yes. The README says it is free and open source under the MIT License, and no paid Orca plan is documented. You still pay your own agent providers, because Orca launches each agent's CLI with your credentials.

The supported-agents page lists Claude Code, Codex, OpenCode, Hermes and OpenClaw among many others, and any CLI agent can be launched from a terminal.

No. The first-session docs say it launches the agent's CLI in the worktree and forwards your subscription credentials.

The desktop app is available for macOS, Windows and Linux; iOS and Android companion apps pair with a desktop host, and orca serve runs a headless server on Linux.