Orca ADE: parallel CLI agents, each in its own git worktree

Orca itself is free: the README says Orca is free and open source under the MIT License. It also does not sell model access — the What is Orca page says Orca runs agents you already use — bring your own Claude, Codex, or OpenCode subscription, and its When to use Orca list opens with You already pay for Claude Code, Codex, or Cursor CLI. The GLM page states verbatim: Orca does not include or resell GLM access. The money sits with that agent's subscription or credentials, not with Orca (checked 2026-09-22).

Not inside Orca. The official first-session page says Orca will launch the agent's CLI with the correct working directory and forward your subscription credentials, and the supported page spells out the mechanism: the agent combobox just launches a process in a terminal. Key and endpoint edits therefore happen in that CLI's own config. The one model switch on the Orca side is --model in orchestration, which the docs limit to opaque provider model IDs for Claude, Codex, and Cursor; every other agent goes back to its own config. Field-by-field wording is on /orca-third-party-api-key (checked 2026-09-22).

Two limits stack up. Coverage: the official usage-tracking page names only Claude Code, Codex, Gemini, OpenCode, Kimi Code, and MiniMax — those 6 — and the supported page says deeper hooks, status, usage tracking, and account switching are noted where supported, so coverage is not uniform. Freshness: the same page says numbers update when the agent writes, not in real time. Whether the bar stays accurate after you point an agent at a third-party endpoint such as OpenRouter is not written in the official docs, and this page does not guess for them (checked 2026-09-22).

You can watch and steer, but it is not a second client. The README says Monitor and steer your agents from your phone, and the mobile page bounds that: gives you a read-mostly view of running agents — agent status, recent terminal scrollback, plus replying to a prompt, sleeping a worktree, reviewing source control and switching agent accounts. Same page: Pairing is one-time and the desktop is always the source of truth, and The mobile companion is in beta. Codex rate-limit reset credits can be spent from the companion's Accounts screen (checked 2026-09-22).

The official ssh page names two blockers. 1) Host key verification: Orca checks built-in SSH connections against your effective OpenSSH known_hosts files and keys it saved before; a rebuilt or reprovisioned host with a new fingerprint is rejected before Orca asks for a password, so verify the fingerprint yourself first, and on a known_hosts mismatch it hands you the matching ssh-keygen -R command. 2) Remote build chain: the first connect installs a small relay on the host, which on Linux often compiles node-pty on the spot, so a machine without make, a C++ compiler and python3 still gets files, git and the editor while remote terminals stay broken until you install the tools and reconnect (checked 2026-09-22).

Not on the usage side: the official wording is No API calls, no extra auth — it only reads the bookkeeping each agent writes to local disk. What calls models and costs money is the launched CLI, and the same docs warn twice: For new launches, Orca pre-fills each supported CLI's permission-bypass flag, and A worktree is an isolated checkout, not a security sandbox. The official countermeasure is choosing Manual under Settings → Agents → Agent Permissions. Orca adds no model proxy layer, since the supported page says the agent combobox just launches a process in a terminal (checked 2026-09-22).