Supported CLIs

  • seats[*].cli picks the launch profile; claude-code, codex, cursor, antigravity and grok are available, and team doctor says what the others still need. vendor, model and version spell one seat's model. account names the budget account the seat spends when one lab has two; without it, the seat spends its vendor, and changing either is an edit only the owner makes.
Every launch profile the build ships: the command, the flags the profile adds, the words it reads on the pane, and the sign-in check
CLIBinaryUnattendedReads the screenSign-in check
claude-codeclaude--dangerously-skip-permissionsready at its ❯ prompt box; working on esc to interruptclaude auth status
codexcodex-a never -s danger-full-access --no-alt-screenready at its › prompt; working on esc to interruptcodex login status
cursorcursor-agent--force --sandbox disabledready at its → prompt; working on ctrl+c to stopcursor-agent status
antigravityagy--dangerously-skip-permissionsready at its > prompt; working on Generating...agy models
grokgrok--always-approveready at its boxed ❯; working on [stop]grok models
geminigemini--approval-mode yoloready at its * or > prompt in the box; working on esc to cancelnone — the hint is gemini
musemuse--yoloready at its ❯ prompt; working on esc to interruptnone — the hint is muse login
julesjulesnoneits session table, read by a screen module; herdr’s status is unreliable, so an idle table counts as freenone — the hint is jules login

The format bullet above names the profiles available today; the three the table adds beyond them — gemini, muse and jules — are the others team doctor reports on.

A pane's screen is read by the profile's own expressions — the words a CLI paints for its idle prompt, its working frame, its dialogs, its questions and the vendor notices it was captured on — and the reading and the live-agent check are separate layers.

A screen that reproduces a CLI's complete idle frame (input row, status row and workspace line, in place) reads idle; the live-agent check is the second layer.

--login reads the logins alone — one line per CLI the file's seats use, said once — and no herdr, session or machine:

Terminal
 team doctor --login ; echo "exit $?"
ok    claude-code: logged in
team doctor: nothing missing, 0 warnings
exit 0

Codex

The Codex profile is tested with CLI 0.157.0 to 0.161.0. Its status line is read for a weekly figure (weekly N% left) when the pane is wide enough to show the number; a cut line is not a figure. It adds -a never -s danger-full-access for unattended execution, plus --no-daemon --no-alt-screen for the captured pane mode, and checks login with codex login status. Rules go as a first message only at an empty idle prompt; delivery is recorded after Codex starts working with the input empty again. Nothing writes a lab's config or an AGENTS.md. /exit is sent only to a free seat. Update and workspace-trust screens are reported and closed without input; the owner handles them before relaunching. Other unrecognised layouts stay unknown.

Antigravity

The Antigravity profile is tested with CLI 1.2.16 to 1.4.2 (agy). It adds --dangerously-skip-permissions for unattended execution, and checks authentication with agy models. Rules go as a first message only at an empty idle prompt; delivery is recorded after the CLI starts working with the composer empty again. Nothing writes a lab's config or an AGENTS.md. /exit is sent only to a free seat. Workspace-trust screens are reported and closed without input; the owner trusts the folder before relaunching. Other unrecognised layouts stay unknown.

The Antigravity profile sets its model with --model.

Cursor

Launching a Cursor seat, like launching cursor-agent by hand, creates Cursor's own project record under ~/.cursor/projects for that folder; team writes no trust (.workspace-trusted) and no Cursor config.

Grok

The Grok profile is tested with CLI 1.0.50 to 1.0.52, runs unattended with --always-approve, rides its rules on the launch line as --rules, and sets its model with --model. It reads the running model off the composer’s footer line — the model’s name, Grok Fast for the fast one, and the version the footer carries — and a screen that names no model, as behind a question widget, falls back to the launch line’s own model, labelled. Grok’s models also run as Cursor seats: Cursor’s own model list carries the Grok model — grok-4.7-high is Grok 4.7 — and Cursor reads the running model off its own status row.

Gemini

The Gemini profile is tested with CLI 0.64.0, runs unattended with --approval-mode yolo, and takes its rules as the seat’s first message. It has no non-interactive login check, so team doctor --login notes gemini: the login is not checked in this version, and the profile’s sign-in hint is gemini.

Muse

The Muse profile is tested with CLI 1.4.9 to 1.4.11, runs unattended with --yolo, and takes its rules as the seat’s first message. It has no non-interactive login check — team doctor --login notes muse: the login is not checked in this version — its Meta account shows only inside the CLI’s own /status, and the sign-in hint is muse login.

Jules

The Jules profile is tested with CLI 0.1.42. Its TUI has no auto-approve flag, so a Jules seat runs unattended through the profile’s own screen walk, not a flag; its rules ride each brief as rules: prepend, and a standing rule — Never delete a Jules session. — rides every brief too, because the TUI’s delete removes the session for the whole account. herdr reports a Jules pane unreliably, so the screen module is the reading: an idle session table counts as free. The jules binary answers no --version, so team doctor reads its version unread. The sign-in hint is jules login.

Other labs through Claude Code

Any Anthropic-compatible endpoint runs as a Claude Code seat — DeepSeek, Z.ai GLM and Kimi Coding are examples. The seat’s launch: names a small launcher script on the PATH instead of the CLI: it points the CLI at the lab’s base URL, reads that account’s API key from the OS keychain — never from the team file — and hands the CLI the arguments it was given. One line is the whole difference, beside cli: claude-code: launch: my-lab-claude. The team file’s own example carries this shape, and model_from: launcher says the launcher picks the model — nothing on the screen can, since Claude Code reads only Opus, Sonnet, Haiku and Fable off its own screen: the seat’s label on team status comes from the file’s display, not the screen.

Not seat kinds yet

The native Kimi CLI is not a seat kind in this build: it has no launch profile, so it cannot be a seat. Kimi’s models run here as Claude Code seats instead, through an Anthropic-compatible endpoint.