Supported agents
Farhelm runs any coding agent in a session, but how much it adds on top depends on the agent. The matrix below is the big picture: what Farhelm does for each agent it knows, and for anything else you run as a custom command. Claude and Codex get the most; the others get less, in different ways. Each agent also has its own page, linked from its name in the matrix, with setup steps and gotchas specific to that agent when it has any.
Some things work the same for every agent and every command, so they have no row: the session keeps running when you close the window, clone and replace with carry its setup over, pasting screenshots and files works, and so do downloads from the terminal.
Container sessions support the same conversation and Restart features as an
ordinary host. Agents in container and sandbox sessions can make $farhelm requests only within their own
container or sandbox (Agents in sandboxes and containers).
| Codex | Claude | Goose | OMP | Pi | Muse | Grok | Cursor | OpenCode | Command | |
|---|---|---|---|---|---|---|---|---|---|---|
| Launch choices | ||||||||||
| Start from the launcher | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ◐ |
| Choose the model | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | – | ✓ | ✓ | – |
| Choose the effort | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | – | – | – | – |
| Guarded or YOLO | ✓ | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ◐ | ◐ | ◐ |
| Workspace trust | ◐ | ◐ | ✓ | ✓ | ◐ | ◐ | ✓ | – | ✓ | – |
| Status | ||||||||||
| Shows working or idle | ✓ | ✓ | ◐ | ◐ | ◐ | ✓ | ◐ | ◐ | ◐ | ◐ |
| Shows when it waits for you | ✓ | ✓ | – | – | – | ◐ | – | – | – | – |
| Conversation | ||||||||||
| Restart resumes the conversation | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ◐ | – | – | ◐ |
| Warns when it cannot resume | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ | ◐ | – | – | ◐ |
| Understands $farhelm requests | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | – | – | – | – |
✓ yes · ◐ partly · – no. Hover a feature name (marked i) or an underlined cell, or tab to it, for the details.
Permission modes
Section titled “Permission modes”The session launcher’s Agent tab also accepts a Wrapper prefix for every agent here. Farhelm keeps composing the agent’s command and runs it through those words; see Agent wrappers for the wrapper’s requirements. The prefix does not change which permission mode you chose.
Choose Custom to supply your own permission flags for Claude, Codex, Grok, Muse, Cursor or OMP. They replace Farhelm’s permission flags; you are responsible for making sure they only change permissions. Farhelm does not check what they do, so the session gets the amber YOLO shield and hosts that ask before YOLO launches still ask.
The field accepts quotes and backslash escapes to group words, without expanding variables or running commands. Its preview shows the words passed to the agent. Templates save them; Clone, Replace with and Restart with let you edit them, while a plain Restart reuses them. Custom is never remembered as the next launch’s default permission. Goose, Pi and OpenCode do not offer this choice. See Start a session.
Guarded lets an agent approve work on its own through a safety check that can still refuse an action or ask you. It does not mean it asks about everything. YOLO skips those checks. The shield beside a session tells you which choice Farhelm started it with; its hover names the agent’s mode and the command it ran.
| Agent | Guarded mode | Flags Farhelm passes |
|---|---|---|
| Claude | Auto mode: its classifier approves or blocks actions | --permission-mode auto |
| Codex | Sandboxed auto: workspace commands run in a sandbox; requests to leave it follow your approval policy | -a on-request -s workspace-write |
| Grok | Auto-review: its classifier approves or blocks actions | --permission-mode auto |
| Muse | Auto-review: its AI reviewer checks eligible actions, with a workspace sandbox | --permission-profile :auto-review |
| OMP | Write: workspace edits are approved automatically; other actions ask you | --approval-mode write |
| Cursor | Agent’s own setting: Farhelm cannot override its configured Run Everything mode | No forcing permission flag |
Guarded is the default for every agent that offers it, unless you chose another mode or the launcher remembered one. Goose keeps its approve, smart approve and chat choices and starts in YOLO unless you choose another mode; Pi and OpenCode offer YOLO only. Custom commands use your own assertion about whether the command is YOLO.
Farhelm passes these flags, but cannot guarantee the agent’s behavior. Claude and Grok can allow actions through your allow rules or hooks; Muse can loosen its sandbox or approval rules; OMP accepts project rules and plugins. For every agent, always-allow answers, managed settings, plugins and switching modes inside the session can change what is allowed, and the agent may change between versions. Check its own mode when that matters. A green shield describes the launch choice, not a live audit of its actions.
For Claude, Grok, Muse and Cursor, Farhelm also tries to notice a no-checks mode on the agent’s status line and changes the shield to YOLO when it sees one. Hover says which mode was seen. This is best effort: detection can be missed or late, and it still cannot see allow rules, hooks or managed settings. Once noticed, the YOLO shield stays for the rest of that agent run, even if you switch back to asking first. This avoids showing a safer shield if a later switch to YOLO goes unnoticed. Restart, Replace or resume starts from the launch choice, as does restarting the host’s supervisor. After a supervisor restart, the shield shows the launch choice until Farhelm sees the status line again. The shield never promises more protection than the launch choice. This changes only the display; Farhelm’s confirmations and permission choices for new sessions still follow the launch.
Cursor’s Agent’s own setting has a grey dashed shield. Hosts that ask before YOLO launches treat it as YOLO, because Farhelm cannot promise a safety check. The same applies to older Claude, Codex, Grok and Muse sessions started without permission flags: restart, plain Replace and copies keep their old command. Use Clone, Replace with or Restart with to build a new guarded launch. On a host that asks first, plain Replace of an older session asks for confirmation too.
Guarded launches need Farhelm 0.29.0 or later on the host. If a remote host runs an older release, update it before starting sessions there.