Which coding agent

A cell runs one agent, chosen in the Agent Picker at the top of an empty cell. Seven agents are first-class, plus Shell, which is not an agent at all.

They are not interchangeable. Each keeps its conversations in its own place, so only the agent that wrote a conversation can continue it — switching the picker changes which “or resume here” list you see. And they reach the GUI tools three different ways, which is the part most worth reading before you pick one.

The seven agents at a glance

	Agent Picker	Command	Badge	GUI tools reach it by	Model override
Claude Code	Claude (default)	`claude`	—	a per-session URL	the Model picker, or a `providers` entry
Codex	Codex	`codex`	`cx`	a per-session URL	`CODEX_MODEL`
Antigravity	Antigravity	`agy`	`agy`	a file in the directory	`ANTIGRAVITY_MODEL`
Grok	Grok	`grok`	`gk`	a file in the directory	`GROK_MODEL`
Muse	Muse	`muse`	`mu`	a plugin, per machine	`MUSE_MODEL`
GitHub Copilot CLI	Copilot	`copilot`	`cp`	a per-session URL	`COPILOT_MODEL`
Cursor CLI	Cursor	`cursor-agent`	`cu`	a file in the directory (not written for you yet)	`CURSOR_MODEL`

Every command can be pointed elsewhere with CLAUDE_BIN / CODEX_BIN / ANTIGRAVITY_BIN / GROK_BIN / MUSE_BIN / COPILOT_BIN / CURSOR_BIN — a pinned version, a wrapper, a path outside PATH.

Which ones tell you when they finish. Claude and Cursor drive both the working dot and the finished one, so a cell you are not looking at raises the attention mark and plays the sound. Codex and Copilot drive the working half only. Antigravity, Grok and Muse drive neither — their cells run fine and simply stay quiet. Claude, and Codex since 6.2.0, can also tell you they are blocked waiting for you — Codex once you trust its hook, the first time a Codex cell asks (6.2.0 guide). For every other agent an approval prompt sits in the cell without a sound.

Nothing has to be installed for an agent you do not use. An agent whose command is missing is dimmed in the new-cell Agent Picker. Pick it and the form says why it cannot start, links the agent’s official install guide, and keeps Start disabled; the others are unaffected. The check runs once when MulmoTerminal starts, so restart MulmoTerminal after installing one. The same guide is named when npx mulmoterminal refuses to start for a missing default agent.

How each one reaches the GUI tools

This is the real difference between them. “GUI tools” are MulmoTerminal’s own MCP tools — the ones that draw a chart in the Canvas, read the workspace’s data, and so on, grouped as render / data / media / external.

There are three routes, and which one an agent takes is a property of the agent’s CLI, not a setting you can change.

1. A per-session URL — Claude Code, Codex and Copilot

In the workspace, these three are handed every tool on one generated URL, per session. There is nothing to register and nothing to switch on: the launcher form does not even show the tool-group toggles there, because they would not add anything.

In a project directory they are handed only the groups that directory registered — still on a per-spawn flag, not by reading the file route 2 uses.

2. A file in the directory — Antigravity, Grok and Cursor

None of these can be handed a URL at spawn, so each reads a config file in the directory and gets whatever that file registers — in the workspace too.

Antigravity reads a JSON file MulmoTerminal writes from the directory’s toggles, and rewrites whenever a toggle flips. Servers it did not write are left alone, and the file is kept out of git status.
Grok reads .grok/config.toml, which is yours — so MulmoTerminal drives grok mcp add rather than editing the file itself.
Cursor reads .cursor/mcp.json, which MulmoTerminal writes from the directory’s toggles — when a cursor cell starts there, not when you flip the switch. (Antigravity’s file is rewritten on the flip; cursor’s is not, so a directory gets the file the first time you actually run cursor in it.) Either way the change reaches the NEXT session, never one already running. Cursor then needs one step neither of the others does: it will not load a server it has not approved, and an unapproved one is silently absent rather than prompted for — so each entry MulmoTerminal wrote is approved through cursor-agent mcp enable as the cell starts. Your own entries in that file are left exactly as you wrote them, and the file is kept out of git status only when MulmoTerminal created it.

So with Antigravity, Grok or Cursor picked, the four toggles stay visible in the launcher form even in the workspace. That is the truthful answer: the directory’s file is the only way these three get any GUI tools.

3. A plugin, per machine — Muse (new in 4.7.0)

Muse has no way to take a URL and no per-directory config either. Its MCP servers are declared by an installed plugin, and muse plugins install records one per machine — installing from one directory does not register anything in that directory.

So MulmoTerminal registers one plugin holding all four group servers, and narrows each session back to what its directory switched on. The plugin is installed and approved through Muse’s own CLI, and an unchanged plugin spawns no subprocess at all.

A group the session is not entitled to serves an empty toolset rather than an error — otherwise a cell that switched one group on would show three broken servers.

Two consequences worth knowing:

The plugin is machine-wide. It appears in muse plugins list for every Muse session on the machine, including ones MulmoTerminal did not start. Those resolve to no session and serve nothing. Remove it with muse plugins remove mulmoterminal; the next Muse cell re-registers it.
Muse’s plugin support is behind an experimental flag of Muse’s own. If a future Muse build renames or drops it, a Muse cell still starts — it just has no GUI tools, with one warning.

Resuming a conversation

OR RESUME HERE lists conversations in the current directory for the agent the picker has selected. Each agent keeps its own store, so the lists never mix.

	Where the conversation lives
Claude Code	its own transcript directory
Codex	a rollout file it writes per session
Antigravity	its own conversation store
Grok	its own store, keyed by directory
Muse	a SQLite session index plus a session log
GitHub Copilot CLI	`~/.copilot/session-state/<id>/`, with a machine-wide SQLite index
Cursor CLI	`~/.cursor/projects/<slug>/agent-transcripts/<id>/`

A resumed Muse session keeps its --workspace, which is what registers its workspace tools — a resume that dropped it came back with the conversation and without the tools (fixed in 4.7.0).

A Muse resume takes no seed prompt. Muse does not accept a prompt on a resume command line, so a seed is only sent on a fresh session.

The header badges

A non-Claude cell wears a short badge (cx, agy, gk, mu, cp, cu) so you can tell at a glance what a cell is running. Beside it the header shows the model and how full the context is, and the up/down arrows are the session’s token usage.

For Muse the context reading is the last completed call, not the largest ever seen — a high-water mark never came down after a compaction and told you to /compact when you need not (fixed in 4.7.0).

Shell is not an agent

Shell starts your OS default shell ($SHELL). Nothing to install, nothing to configure, no conversation and no GUI tools.

The launch commands below the picker are the same idea: they run the command line you wrote, verbatim. MulmoTerminal does not read that command — so a launch command whose text is claude is a terminal with Claude Code in it, not an agent session: no session id, no resume, no GUI tools, no badge. The Agent Picker is what starts an agent.

Running Claude Code differently

Two things widen what “Claude” means, and both are Claude-only.

A different backend or model — `providers`

A providers entry in ~/.mulmoterminal/config.json registers any Anthropic-compatible backend — OpenRouter, Moonshot, a local Ollama bridge, a company gateway — and it then appears in the Model picker beside Anthropic’s own models. A directory can pin its own provider / model so a project always runs on the same one.

Full setup, the measured pass rates of the built-in model list, and the misconfigurations that are hard to diagnose from inside a session: Providers and models.

Your own command line — `customAgents`

A customAgents entry is your way of starting Claude Code — a wrapper script, a pinned binary, ollama launch claude --model … -- — and it appears in the Agent Picker beside the built-in ones. Claude Code’s whole argv is appended to what you wrote, so the session still resumes, still reports cost, and still gets the GUI tools.

This is what separates it from a launch command: the entry declares agent: "claude", so MulmoTerminal knows which CLI’s arguments to append. Only Claude is supported — an entry for another agent is not a label that would work.

See Configuration for the entry shape.

Starting without Claude Code

npx mulmoterminal refuses to start when it cannot find claude. That is deliberate — Claude Code is what almost every installation runs, and starting silently on some other agent would leave you guessing which one answered. But if you work in Codex or Copilot and have never installed Claude Code, that gate is the whole app (#2082).

Declare a default agent and the check follows your declaration instead. Either of these works:

npx mulmoterminal --agent codex

~/.mulmoterminal/config.json:

{ "defaultAgent": "codex" }

The flag wins over the file, and the flag is not written back — it is a statement about one launch. Valid names are the seven agent ids: claude, codex, antigravity, grok, muse, copilot, cursor.

Declaring one changes exactly two things:

Start-up checks that agent instead of Claude Code. It is still required — you named it, so a machine without it cannot do what you asked, and being told beats an empty grid. A missing one names the command it looked for and the <AGENT>_BIN variable that overrides it.
A new cell opens on that agent. The Agent Picker starts there rather than on Claude.

What it does NOT change

Cells you have already saved keep the agent they were saved with. A grid cell records its agent only when it is not Claude, so a Claude cell is stored as the absence of that field — which means “Claude”, permanently, however you later set defaultAgent. Changing this setting never re-points an existing cell, and a browser that has already used the Agent Picker keeps the choice it remembers.

That is why this is called the default agent rather than the agent: it seeds new things.

`CLAUDE_BIN` is honoured

If you keep Claude Code somewhere that is not on PATH, set CLAUDE_BIN to its full path and start-up finds it. The start-up check used to look for the literal word claude on PATH while the server ran CLAUDE_BIN, so a perfectly good install could be refused; that is fixed. Every agent has the same override — CODEX_BIN, GROK_BIN, CURSOR_BIN and so on.