moai cc / cg / glm Launchers

moai cc, moai cg, and moai glm are three launchers that run Claude Code with different backend configurations. All three adjust settings and then exec to replace the current process with Claude Code. Since which model does which work is what drives cost, the launcher choice is the first tokenomics decision.

moai cc [-p profile] [-w [name]] [-- claude-args...]

Removes GLM-specific environment variables from .claude/settings.local.json, resets team mode if it was enabled, and then runs Claude Code.

Info

-k is the kanban chain token, and -f is the dedicated entry token for Factory Mode. One -k is still read three ways — no argument or a SPEC-ID is the kanban lead, --name <role> is a kanban companion, and a number is a lane run. A launch takes one entry token only, so -k together with -f is an error. The mixed-backend launcher moai cg refuses both modes (the factory-side refusal sentinel is FACTORY_MODE_UNSUPPORTED_BACKEND). For the full contract see Kanban Mode and manager-lead Lead Coordinator.

A card travels differently here than in kanban. In kanban one card moves across the plan → run → sync columns, while in a factory one card goes whole to one lane and passes the three phases serially inside it. Each phase is spawned by that session as Agent() subagents, and write-capable spawns are isolated with isolation: "worktree". A lane runs at most 10 concurrent subagents, and the launcher seeds that value into lanes and companion sessions through CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, so N lanes sharing one machine’s capacity is guaranteed by configuration rather than by operator restraint. Never bring the lanes up all at once — start the first, confirm it is actually producing output, then activate the rest.

The backend mix is decided by token availability first. One starting point is the lead on GLM, plan on Claude (Opus), run on GLM, and sync on Claude (Opus), placing Opus only on the phases where judgment is heavy. A different combination, or unifying on a single backend, is equally fine.

The permission mode is one of default, acceptEdits (project default), plan, auto, bypassPermissions, dontAsk. The auto mode runs a background classifier that inspects actions and requires a Team plan + Sonnet/Opus 4.6 or later.

moai glm setup <api-key>   # Save API key (first time only)
moai glm                   # Run with the GLM backend
moai glm -p work           # Run with the 'work' profile
moai glm status            # Check credential status

Reads GLM credentials from ~/.moai/.env.glm, injects environment variables such as ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL, and then runs Claude Code.

Warning

GLM does not support the auto permission mode (it is a third-party provider). If you need auto, use moai cc or moai cg. Also, Z.AI has a low concurrent-request limit (1-3 in-flight on the paid tier), so for parallel multi-agent execution the moai cg hybrid mode is more stable.

moai cg [-p profile]

CG stands for “Claude + GLM”, a cost-optimized team configuration.

• Leader (current tmux pane): uses Claude models (opus/sonnet)
• Teammates (new tmux panes): use GLM models via the Z.AI proxy

On launch it validates the tmux session, removes the GLM environment in the leader pane (Claude), injects the GLM environment into the tmux session (teammates), and sets teammateMode=tmux and team_mode: cg.

Prerequisites:

1. Set the GLM API key with moai glm setup <api-key>
2. Run inside a tmux session for per-pane environment isolation

All three launchers accept -p <name> to select a named profile, which sets CLAUDE_CONFIG_DIR to ~/.moai/claude-profiles/<name>/. Use this to keep multiple accounts / setting sets separate.

All three launchers accept -w [name] to start the session inside an isolated git worktree, collapsing the two-step cd then launch into a single command.

moai cc -w feat-login    # Start in .claude/worktrees/feat-login/
moai cc -w               # Auto-generated name
moai glm -w feat-login   # Same for the GLM backend
moai cg -w feat-login    # Same for the hybrid

Behavior:

• The worktree path is .claude/worktrees/<name>/. <name> is a worktree name — not a branch name and not a SPEC ID.
• If a worktree of that name already exists it is reused rather than recreated, so this doubles as the re-entry path into a tree an earlier session was working in.
• Omitting the name lets Claude Code generate one.
• The -w=name, --worktree name, and --worktree=name spellings are all accepted and mean the same thing.
• Arguments after -- pass through to Claude Code untouched and are unaffected by this rewrite.

Info

Naming the worktree after the SPEC ID (moai cc -w SPEC-XXX-001) lets a session handoff bring the next session back into the same working tree with one line.

• CG Mode (Claude + GLM)
• Profile Management
• Security Notes — GLM credential path security model
• CLI Overview