Harness support

Muster keeps routing, scoring, manifests, waves, and gates in one deterministic Node CLI. The model-facing controls depend on the active harness. Native Claude Code, Codex, Kimi, and Cowork lanes use that harness's account or subscription; Muster does not ask those lanes for a separate model API key. ChatGPT Work is a private/local, proof-gated exception that uses Secure MCP Tunnel and a separately billed OpenAI Platform API key.

Claude Code, Codex, and Kimi expose all ten modes, including Design and Init. Cowork and ChatGPT Work use narrower MCP surfaces whose verified boundaries are listed below.

Support matrix

Harness	Entry points	Parallel dispatch	Write isolation	Policy enforcement	Init handoff
Claude Code	Ten /muster:* modes, including Design	Native agents; optional Workflow fan-out	Native agent worktrees	Plugin hooks plus review gates	Native /init, canonical instruction pair
Codex	Ten $muster-* modes, including Design	Production waves use process-only codex-wave / codex exec -C	Authenticated pristine registered worktrees	Advisory hooks plus review gates	Native /init, canonical instruction pair
Kimi	Ten /muster-* modes, including Design	In-session native subagents; attended process lane is report-only	Muster-created worktrees and base-SHA receipts	Native permission rules plus review gates	Unavailable; explicit acknowledgement required
Cowork	MCP instructions and tools	Probe phase 3 required; sequential path is verified	No proven per-subagent worktree primitive	Review gates; native plugin enforcement is unverified	No proven callable native adapter
ChatGPT Work (private/local; proof-gated)	Universal plugin + registered MCP connection	Pro-safe muster_prioritize only after native Scan Tools; full profile requires entitlement + double opt-in	Host-controlled; no Muster worktree claim	ChatGPT approvals/admin controls plus nonce-bound native proof gate	No callable Init adapter; no Codex-config inheritance

On a narrow screen, focus the comparison and scroll sideways; the harness name stays pinned. Prefer a harness guide below when you need the full contract without horizontal comparison.

The matrix separates verified behavior from declared capability. A file present on disk does not prove that a harness loaded it. Codex and Kimi installers can inspect their owned files; Cowork's native plugin lane remains opt-in through MUSTER_COWORK_NATIVE_PLUGIN because the local MCP server cannot probe the host's plugin loader.

For Claude Code and Codex Init, AGENTS.md is authoritative, and CLAUDE.md contains exactly:

# Claude Code

@AGENTS.md

If conflicting instruction files existed at the preparation baseline, Init leaves a HUMAN-HOLD instead of overwriting or merging them.

For ChatGPT Work, the local/repo Plugins Directory source is documented for the desktop proof lane: restart or refresh ChatGPT Desktop, select the source, and install the plugin. Do not infer that the same local source is ingested by Work web; use web only with an independently supported source. Work remains separate from Codex configuration and inheritance. A configured connection or successful tool scan does not prove native invocation; the active host must produce the completed Muster card and matching nonce-bound server evidence.

Codex Desktop on native Windows is also distinct from both ChatGPT Work and a Codex installation inside WSL. Native Windows and WSL normally resolve different homes, Node runtimes, plugin caches, and CODEX_HOME directories. Install and diagnose Muster in the same host that launches Codex.

ChatGPT Desktop surface detection

ChatGPT Desktop is a host shell, while Codex Desktop and ChatGPT Work (also called GPT Work) are selected experiences with different capabilities. A deterministic process cannot inspect the app's current mode, so Muster requires an explicit declaration and fails closed:

• muster desktop-harness chatgpt-desktop reports that an experience must be selected. It exposes no capability lane and no native init command.
• muster desktop-harness codex-desktop selects capabilities --codex. muster init uses a HUMAN-HOLD for Codex's native /init, expecting AGENTS.md; the invocation alone is not completion evidence.
• muster desktop-harness gpt-work selects capabilities --work, whose dispatch floor is registered MCP tools or inline execution. Work has no proven native instruction initializer and does not inherit Codex /init or AGENTS.md, so muster init records an unavailable HUMAN-HOLD with no expected artifact until the user explicitly acknowledges it.

Choose a guide

• Codex covers scoped profiles, hooks, trust, and receipts.
• Kimi covers native agents, skills, permission rules, and model lanes.
• Cowork covers the local MCP server, its 31-tool surface, and the verified degradation path.
• ChatGPT Work covers the private/local universal plugin, Secure MCP Tunnel, profile opt-ins, and the native proof contract.
• Install covers Claude Code.

Configuration shared by all lanes lives in Configuration. Harness-specific settings are grouped there instead of being repeated as unofficial knobs.