Features and options

Each feature is configured under features.<name> in ~/.mu/agent/mu.json. false turns a feature off, true turns it on with its defaults, and an object overrides single options:

{
  "features": {
    "injection": { "tools": ["web_fetch", "mcp__*"] },
    "board": { "defaultOn": true },
    "ttsr": false
  }
}

The desktop app's settings write the same file, with the same names and descriptions as this page. Both come from one manifest in the repository.

Message preflight `preflight`

On by defaultDecision points input.preflight

Judges a message before work starts and shows the verdict in the conversation.

Option	Default	Meaning
`thinking`	`false`	Set the thinking level from the verdict (a switch loses the cache)
`hints`	`true`	Give the model a one-line hint
`show`	`true`	Show the wait and the verdict
`waitMs`	`4000` ms	Wait at most. After this the turn starts without the verdict. Range 500 to 30000.

Mid-run messages `interjection`

On by defaultDecision points input.interjection

Handles messages sent while the agent is working.

Option	Default	Meaning
`waitMs`	`1500` ms	Wait at most. Range 200 to 10000.

Skill disclosure `skills`

On by defaultDecision points skills.disclosure

Chooses which skill descriptions a session carries.

Option	Default	Meaning
`minSkills`	`4`	Do not filter below this many skills. Range 1 to 100.
`waitMs`	`4000` ms	Wait at most. Range 500 to 30000.

Capability catalog `catalog`

On by defaultDecision points capability.disclosure

Install a lot, expose little: packs and MCP servers stay hidden and open per task.

Option	Default	Meaning
`waitMs`	`4000` ms	Wait at most. Range 500 to 30000.

Task frame and to-do list `frame`

On by defaultDecision points task.frame

Keeps what you actually want: the goal, your hard constraints (your own words, with their source), the current subgoal and the acceptance items. Those items are the to-do list the model ticks with the todo tool. Every goal-relevance judgment reads it, and it rewinds with the session.

Option	Default	Meaning
`waitMs`	`3000` ms	A turn waits for its frame at most. Counted from the arrival of the message. A late update is not dropped: the turn starts on the previous version plus the raw message, and the update lands when it is ready. Range 0 to 30000.
`writerTimeoutMs`	`20000` ms	Time limit for writing a frame. On a timeout or a malformed reply the last version stays and is marked stale, goal-based filtering holds back, and the next message retries. Range 1000 to 120000.
`show`	`true`	Show frame updates in the chat

Lessons `memory`

On by defaultDecision points memory.recall memory.capture memory.outcome memory.worth memory.merge memory.applied

Learns from your corrections, from the agent getting itself unstuck, and from what the model and sub-agents found; merges each lesson with the kept ones before storing it, brings the relevant ones along next time, and retires those nobody follows. /lessons shows them, /forget retires one.

Option	Default	Meaning
`path`	empty	Lessons file. Empty uses the default location.
`maxCandidates`	`24`	Judge at most. The most followed, then the most recently stored or confirmed, come first. Range 1 to 200.
`maxInjected`	`5`	Bring in at most. Range 0 to 20.
`waitMs`	`4000` ms	Wait for the recall at most. Counted from the arrival of the message; picking lessons for sub-agents waits as long. Range 0 to 30000.
`retireAfter`	`8`	Retire after this many recalls never followed. 0 never retires a lesson by itself. Range 0 to 100.
`mergeNeighbours`	`6`	Compare with this many similar lessons. Picked by shared words, asked in one request; 0 stores without merging. Range 0 to 16.
`outcome`	`true`	Learn from the agent getting unstuck. Needs the writer model or the conversation's model to put the lesson into words.
`applied`	`true`	Check at the end of a turn whether lessons were followed

Risky command guard `guard`

On by defaultDecision points tool.risk

Confirms dangerous commands before they run.

No options.

Permission modes `permissions`

On by defaultDecision points tool.approval

Full access, Jev approves, or minimal permissions, switched any time with /permissions. When a step needs your permission the status bar asks; allow it once or for the whole conversation. While it is on, it also handles the risky command guard.

Option	Default	Meaning
`mode`	`jev`	Mode of a new conversation. Once you pick one with /permissions, your last pick is used.`full` Full access · `jev` Jev approves · `ask` Minimal permissions

Hard constraint gate `constraints`

On by defaultDecision points tool.constraint

Checks a change or a command against your hard constraints first; sub-agents work under them too.

Option	Default	Meaning
`maxConstraints`	`6`	Constraints checked per call at most. The newest ones are used. Range 1 to 20.
`waitMs`	`5000` ms	Wait at most. Range 500 to 30000.

Injection screening `injection`

On by defaultDecision points tool.injection

Withholds instructions aimed at an AI before the model reads web pages, search results and MCP answers.

Option	Default	Meaning
`tools`	`web_fetch`, `web_search`, `browse`, `mcp__*`	Tools screened. Tool names, or a prefix ending in ; mcp__ is every MCP server.
`waitMs`	`6000` ms	Wait at most. After this only plain injection phrases are withheld, by rule; the verdict is still recorded. Range 500 to 30000.

Tool output admission `admission`

On by defaultDecision points tool.admission tool.admission.test-log

Controls how long tool output enters the context.

Option	Default	Meaning
`minChars`	`4000` chars	Only filter output longer than. Range 500 to 200000.
`chunkChars`	`1200` chars	Chunk size. Range 200 to 20000.
`maxChunks`	`48`	Judge at most this many chunks. Range 4 to 400.
`batchChunks`	`16`	Chunks per request. A hosted judge takes many chunks in one request, the state billed once; the local judge takes one at a time. Range 1 to 32.
`waitMs`	`4000` ms	Wait at most. After this the output goes in whole; the verdict is only recorded. Range 500 to 30000.
`passThrough`	`read`, `edit`, `write`	Never filter these tools
`testLog`	`off`	Test log trimming`off` Off · `rules` Drop exact repeats only (lossless) · `jev` Also let the judge select from the rest
`agentEnv`	`true`	Tell test runners an agent is reading. Vitest and others then print failures and the summary only.

Forgetting stale results `forgetting`

On by defaultDecision points context.forget

Replaces stale results with tombstones when the context gets tight.

Option	Default	Meaning
`thresholds`	`50`, `70`, `85`	Trigger thresholds (context use %)
`minChars`	`6000` chars	Only results longer than. Range 500 to 200000.
`minAgeTurns`	`2`	At least this many turns old. Range 0 to 50.
`maxPerBatch`	`12`	Judge at most per batch. Range 1 to 100.
`waitMs`	`4000` ms	Wait at most. This request waits this long; verdicts that come later apply at the next request. Range 500 to 30000.

Summary-free compaction `compaction`

Off by defaultExperimentDecision points context.compact

Compaction that keeps original passages instead of a summary.

Option	Default	Meaning
`keepThreshold`	`0.5`	Keep threshold. Range 0 to 1.
`minChars`	`600` chars	Keep anything shorter than. Range 100 to 20000.
`headChars`	`300` chars	Head kept when pruning. Range 0 to 5000.
`targetRatio`	`0.5`	Target share of the original. Range 0.05 to 1.
`maxWindowShare`	`0.25`	At most this share of the window. Range 0.05 to 0.9.
`freeChars`	`24000` chars	Histories this small are not squeezed. Range 0 to 500000.
`maxJudged`	`120`	Judge at most this many passages. Range 10 to 1000.

Checkpoints and rewind `checkpoint`

On by defaultDecision points turn.rewind

Before the first change of every turn that edits files, snapshots the working tree into mu's own shadow git directory, never touching your repository. /rewind takes files and conversation back, /rewind undo takes the rewind back.

Option	Default	Meaning
`dir`	empty	Where snapshots are kept. Empty means mu/checkpoints in the mu home, one shadow repository per project.
`keep`	`50`	Snapshots kept per project. Range 1 to 1000.
`maxAgeDays`	`14` days	Days a snapshot is kept. Range 1 to 365.
`maxFileMb`	`5` MB	Largest file in a snapshot. Larger files stay out of snapshots, and a rewind never touches them. Range 1 to 1024.
`maxFiles`	`5000` files	Most files a snapshot takes in. With more files to take in, the session goes without checkpoints and says so once. Your home folder and mu's own folders are never snapshotted. Range 100 to 1000000.
`maxTotalMb`	`200` MB	Most a snapshot takes in, in all. With more to take in, the session goes without checkpoints and says so once. Range 1 to 100000.
`ignore`	empty	Extra paths to ignore. gitignore patterns, one per line. The project's own .gitignore and the built-in list (node_modules, build outputs and so on) always apply.
`timeoutMs`	`30000` ms	Longest a snapshot may take. Past this the turn goes without a checkpoint; the tool call still runs. Range 1000 to 600000.
`propose`	`true`	Propose a rewind at a dead end
`confirmSeconds`	`120` s	How long a proposal waits for you. Unanswered, the run carries on. 0 waits forever. Range 0 to 3600.

Drift monitor `monitor`

On by defaultDecision points turn.drift

Notices drift from the goal and going in circles.

Option	Default	Meaning
`every`	`6`	Check every this many tool calls. Range 1 to 100.
`repeats`	`3`	Repeats that count as a loop. Range 2 to 20.
`window`	`8`	Window. Range 2 to 100.

Language server diagnostics `lsp`

On by defaultDecision points diagnostics.delivery

Uses the language servers already installed here (installs none), keeps only what an edit newly introduced, and lets the judge pick the moment to tell the model.

Option	Default	Meaning
`builtin`	`true`	Use the built-in table of well-known servers. Well-known servers for TypeScript, Python, Go, Rust and C/C++, looked up on PATH only. Your own go under features.lsp.servers in mu.json.
`editTools`	`edit`, `write`	Tools that count as editing a file
`subAgents`	`false`	Sub-agents start language servers too. Off by default: one set of servers per sub-agent is heavy on the machine.
`settleMs`	`1500` ms	Wait for the server after an edit, at most. Past this nothing waits: late diagnostics are handled at the next edit or when the turn ends. Range 0 to 10000.
`turnEndSettleMs`	`3000` ms	Wait at the end of a turn, at most. Range 0 to 30000.
`quietMs`	`250` ms	Silence that counts as the server being done. Range 20 to 5000.
`baselineMs`	`5000` ms	Wait for the pre-edit diagnostics, at most. Waited for in the background, never blocking a tool. Past it, the first report is not called new. Range 100 to 60000.
`maxItems`	`10`	Diagnostics told at once, at most. Range 1 to 100.
`maxServers`	`4`	Servers running at once, at most. Range 1 to 16.
`waitMs`	`4000` ms	Wait for the judge, at most. Range 500 to 30000.

Mid-stream correction (experiment) `ttsr`

Off by defaultDecision points output.drift

Semantic TTSR: the judge, not a regular expression, notices the output going astray, cuts it, shows the rule and lets the model carry on. Off by default; when on, every stretch of output costs one more judge call.

Option	Default	Meaning
`rules`	empty	Rules to hold at all times. One per line, in plain words, e.g. “Answer in Chinese”, “No placeholder implementations”. Your hard constraints from the task frame are added automatically.
`segmentChars`	`600`	Characters of output per check. Range 200 to 5000.
`maxRules`	`4`	Rules checked per call at most. When there are more, the ones you said last stay. Range 1 to 12.
`maxInterrupts`	`2`	Cuts at most between two of your messages. Range 1 to 10.

Goal mode `goal`

On by defaultDecision points goal.met

After /goal <condition> the agent keeps working until the condition holds (/goal alone asks for it). Each time it wants to stop, a model reads the evidence: met, or the next step. It stops by itself when you interrupt, a model call fails, it idles or goes in circles, or its allowance runs out; your next message picks it up again.

Option	Default	Meaning
`checker`	`model`	Who checks whether the goal holds. A model reads the goal, what the run did, the last test run and the closing message, and gives a verdict and the next step; Jev only when it gives no answer. Either way an open acceptance item or an unverified edit means not yet.`model` A model (recommended) · `jev` The Jev judge
`checkModel`	empty	Model for the check. As provider/model. Empty: the session's current model.
`checkThinking`	`off`	Thinking level of the check`off` Off · `minimal` Minimal · `low` Low · `medium` Medium · `high` High
`checkTimeoutMs`	`90000` ms	Longest wait for the check. After that, Jev decides, and failing that the facts alone. Range 5000 to 600000.
`stallLimit`	`2`	Runs without progress before it stops. The check found the agent repeating itself. The first time it is told to change course; at this count the goal pauses for you. Range 1 to 10.
`maxContinuations`	`20`	Continuations at most. Counted from your last message. Range 1 to 200.
`maxMinutes`	`180` min	Longest automatic run. Counted from your last message. Range 5 to 1440.
`idleLimit`	`2`	Idle runs before it stops. Runs in a row that ended without a single tool call: the agent is going nowhere. Range 1 to 10.

Plain-language board `board`

On by defaultDecision points board.read

Tells you in plain words how far the work is, what is happening now and what waits on you, and keeps a running account: a line for every step as it happens, and what the agent says or finds retold at once by a plain-speaking model, with Jev deciding what is worth telling; a finished run is summed up once more, with the account kept. For you only, never in the model's context. Switched per project (/board on, /board off); the first time, you pick the model that writes it (/board model changes it).

Option	Default	Meaning
`defaultOn`	`false`	On for projects not switched yet. Off by default: each update of the board costs a model call.
`model`	empty	Model that writes the board. As provider/model; set here, nobody is asked. Empty: you pick one the first time the board is switched on (Claude Opus 4.6 recommended, then Gemini 3.8 Flash), kept in mu/board.json; none picked: the writer model, else the session's.
`language`	`auto`	Language of the board`auto` The one you write in · `zh` Chinese · `en` English
`everyTools`	`5`	Tool calls between two looks. Range 1 to 100.
`minIntervalMs`	`20000` ms	Shortest gap between two looks while it works. When the agent stops it always looks, whatever this says. Range 0 to 600000.
`maxSteps`	`10`	Latest tool calls read each time. Range 3 to 50.
`narrateTimeoutMs`	`60000` ms	Longest wait for the writer. After that, fixed sentences are used. Range 5000 to 300000.

Completion check `completion`

On by defaultDecision points turn.completion

Checks that something verified the work before it is called done.

Option	Default	Meaning
`waitMs`	`3000` ms	Wait at most. After this there is no nudge; the verdict is only recorded. Range 500 to 30000.

Carry on after stopping short `continuation`

On by defaultDecision points turn.continue

Sends a run that stopped short of what it said it would do back to it.

Option	Default	Meaning
`maxNudges`	`2`	Most nudges per message. Range 1 to 5.
`waitMs`	`3000` ms	Wait at most. After this there is no nudge; the verdict is only recorded. Range 500 to 30000.

Notification routing `notify`

On by defaultDecision points notify.routing

Decides when events are told to the model.

Option	Default	Meaning
`budgetThresholds`	`70`, `85`	Context budget notices (%)

Cache warming `warming`

On by defaultDecision points cache.warming

Refreshes the prompt cache before it expires when that pays off.

No options.

Judge connection warm-up `warmup`

On by default

While the session is in use, one tiny question keeps the judge's connection warm: the questions before a turn skip 1-3 s of connecting.

Option	Default	Meaning
`intervalMs`	`50000` ms	Every. The server drops a connection idle for about 90 s; stay under that. Range 10000 to 85000.
`idleMs`	`900000` ms	Stop after this much quiet. Range 60000 to 7200000.

Sub-agents `swarm`

On by defaultDecision points swarm.routing swarm.patch

Delegates tasks to sub-agents with roles.

Option	Default	Meaning
`models`	empty	Model ladder (cheapest to strongest). Empty means every sub-agent uses the session's model.
`maxTasks`	`6`	Tasks per call at most. Range 1 to 32.
`concurrency`	`3`	Running at once. Range 1 to 16.
`defaultAgent`	`worker`	Default role
`agentsDir`	empty	Folder with your own roles. Empty uses <agent dir>/agents.
`isolation`	`worktree`	Sub-agents that edit files. Isolated changes come back as a patch that the main agent applies with apply_patch_from, or not. Outside a git repository it falls back to editing in place.`worktree` Edit in a git worktree of its own · `none` Edit in place
`carryUncommitted`	`true`	Carry uncommitted changes over. The sub-agent starts from the files as the main agent sees them, not from the last commit.
`patchPreviewLines`	`30`	Lines of patch preview. The beginning of the patch shown with the sub-agent's report. Range 0 to 400.

Hive `hive`

On by defaultDecision points hive.publish hive.deliver hive.relate

Several bees work one hard task in parallel; the judge gates what passes between them. /hive <question> starts one yourself; the agent also starts one when a problem resists a direct attempt.

Option	Default	Meaning
`maxNotesPerBee`	`12`	Notes per bee at most. Range 1 to 100.
`maxDeliveriesPerBee`	`10`	Deliveries per bee at most. Range 1 to 100.
`checkpointEvery`	`4`	Ask for findings after this many silent tool calls. 0 turns this off. Range 0 to 50.
`lastCall`	`true`	One last hearing while writing the report
`maxRelatedPerNote`	`6`	Earlier findings a new one is held against, at most. Only those sharing words with it. 0 turns corrections off. Range 0 to 30.
`verifyConflicts`	`1`	Verifier bees for unsettled disputes, at most. 0 adds none; both sides stay in the report. Range 0 to 3.
`verifyAfterSeconds`	`60`	Seconds a dispute may stand before a verifier is added. Range 0 to 600.

File location `locate`

On by defaultDecision points files.locate

The locate tool: find files by description.

Option	Default	Meaning
`candidates`	`40`	Candidates judged. Range 5 to 400.
`results`	`12`	Results returned. Range 1 to 100.

Bulk judging tool `judgeItems`

On by defaultDecision points judge.items

The judge_items tool: one yes/no question, a probability for each of many items.

Option	Default	Meaning
`maxItems`	`500`	Most items per call. Range 10 to 5000.

Built-in browser `browser`

On by defaultDecision points browser.step

A judge-driven browser with its own profile; it never touches yours.

Option	Default	Meaning
`headless`	`true`	Run headless
`maxSteps`	`40`	Steps per run at most. Range 1 to 200.
`textChars`	`4000` chars	Page text read per step. Range 500 to 50000.
`embedded`	`true`	Use the desktop app's browser panel when it is running. Every step is then visible and can be paused, stopped or taken over; without the app mu's own browser is used.
`profileDir`	empty	Browser profile folder. Empty uses mu's own folder.

Background commands `background`

On by default

bg_start / bg_output / bg_stop: dev servers and long builds without blocking the turn. Whether the end of a job interrupts is decided by notification routing; every job stops with the session.

Option	Default	Meaning
`maxJobs`	`8`	Jobs running at the same time. Range 1 to 32.
`bufferChars`	`262144` chars	Output kept in memory per job. The newest is kept; the log file has everything.
`maxOutputChars`	`20000` chars	Most that one bg_output returns
`killGraceMs`	`3000` ms	Wait before a job is killed by force. SIGTERM first (taskkill without /F on Windows), force after this long.
`logDir`	empty	Log directory. Empty means ~/.mu/jobs/<session id>, cleared after 7 days.
`inheritShell`	`true`	Use the foreground shell settings. Reads shellPath and shellCommandPrefix from settings.json, like the bash tool.
`wakeWhenIdle`	`false`	A finished job may wake an idle agent. Only a verdict of "now" starts a turn; when off the notice is only appended.

Web reading and search `web`

On by default

web_fetch reads a page as text (time and size limits, internal addresses refused, page text labelled untrusted); web_search uses a source that answers from mainland China without a key.

Option	Default	Meaning
`search`	`so`	Search source. so (360), sogou, bing-cn, searxng:<base url>, or a URL with {query} in it.
`searchFallback`	`true`	Try the other built-in sources on failure. Applies to a verification page, no results, or results unrelated to the query.
`maxChars`	`20000` chars	Most that one web_fetch returns
`timeoutMs`	`20000` ms	Time limit of one request
`maxBytes`	`2000000` bytes	Download limit
`maxRedirects`	`5`	Redirects followed. Range 0 to 20.
`allowLoopback`	`true`	Allow this machine's dev servers. Only when the URL itself says localhost or 127.0.0.1; a redirect from the web is always refused.
`allowPrivate`	`false`	Allow private network addresses. 10/8, 172.16/12, 192.168/16, link-local and the like. Refused by default so a page cannot steer the reader into the local network.

Welcome screen `welcome`

On by default

The welcome box in the terminal and the identity note.

No options.

Inherit existing setup `inherit`

On by default

Uses the rules, skills and MCP servers you already set up for Claude Code, Cursor and Codex, read-only. Always-on rules join the prompt; a rule scoped to file patterns is handed over once, with the result of the first tool that touches a matching file; the rest are listed by description. Project content is only used for a trusted project.

Option	Default	Meaning
`claude`	`true`	From Claude Code
`cursor`	`true`	From Cursor
`codex`	`true`	From Codex
`rules`	`true`	Inherit rules
`skills`	`true`	Inherit skills
`mcp`	`true`	Inherit MCP servers. When off, only the servers under mcp.servers in mu.json are used.
`maxRuleChars`	`4000` chars	Most of one file-scoped rule handed over. The rest stays in the file for the model to read. Range 500 to 20000.
`maxAlwaysChars`	`16000` chars	Budget for always-on rules. Always-on rules that do not fit are listed by description instead, so rules cannot bloat the prompt. Range 0 to 100000.

MCP servers `mcp`

On by default

MCP servers are connected by pi's own client (stdio and Streamable HTTP, OAuth sign-in, managed with /mcp). Each server mu takes over from Claude Code, Cursor, Codex and mu.json is a catalog entry that stays hidden: it starts and its tools are handed to the model only when the judge finds the task needs it or the model asks through find_capability. Tool lists are cached, so the judge has a description before a server ever ran. A server defined by a project asks before its first start, and again when its definition changes. "exposure": "always" under mcp.servers in mu.json keeps a server open. Servers added with mu mcp add are pi's: they connect with the session, and tool_search loads their tools when needed.

Option	Default	Meaning
`startTimeoutMs`	`45000` ms	Start timeout. An npx-style server downloads itself first. Range 1000 to 300000.
`requestTimeoutMs`	`120000` ms	Call timeout. Range 1000 to 3600000.

Capability packs `packs`

On by defaultDecision points review.triage

Tool packs that ship with mu and stay hidden until a task needs them; a missing program gives an install hint.

Option	Default	Meaning
`astGrep`	`true`	ast-grep structural search and rewrite. Finds and changes code by syntax, not text. Needs ast-grep on this machine.
`maxResults`	`50`	Matches one search returns. Range 1 to 500.
`maxDiffChars`	`12000` chars	Most diff one result carries. Range 1000 to 200000.
`astGrepCommand`	empty	Path of ast-grep. Empty: ast-grep, then sg, from PATH.
`github`	`true`	GitHub through gh. No tools, one skill that teaches gh for PRs, issues, checks and releases. Needs gh on this machine.
`ghCommand`	empty	Path of gh. Empty: gh from PATH.
`commit`	`true`	/commit: split the change into commits. A model proposes the groups and the messages; commits are made only after you confirm the plan, all or none. Never pushes.
`maxPlanChars`	`60000`	Characters of the change the model reads. Range 5000 to 400000.
`review`	`true`	/review, with findings sorted P0 to P3
`conflicts`	`true`	Conflict resolution. When a merge, rebase or cherry-pick stops on conflicts: both sides and their base per file, the resolution written block by block and the file marked resolved. Never commits, never runs --continue.
`maxSideLines`	`80`	Lines shown per side of a conflict block. Range 10 to 1000.
`maxConflictChars`	`20000`	Characters of one file's conflicts shown. Range 2000 to 200000.
`maxFindings`	`40`	Findings sorted per triage at most. The rest are still reported, just not sorted. Range 5 to 200.
`debugger`	`true`	Debugger: breakpoints, stepping, variables. Runs a program under a debugger, stopped at breakpoints or an uncaught exception, with the call stack and variables. debugpy for Python, delve for Go, lldb-dap for compiled programs, each if installed; add your own under debugAdapters in mu.json. One run at a time, ended with the session.
`maxFrames`	`20`	Frames shown at a stop. Range 1 to 200.
`maxVariables`	`50`	Variables listed per scope. Range 5 to 1000.
`debugOutputChars`	`4000` chars	Program output one result carries. Only the tail is kept when there is more. Range 500 to 100000.
`debugWaitMs`	`30000` ms	Longest wait for the program to stop. If it has neither stopped nor ended by then, the model is told it still runs, and can pause it to see where, or wait more. Range 1000 to 600000.

Sign in with Google (experimental) `googleLogin`

On by defaultExperiment

Adds two entries to /login: Gemini CLI's sign-in for the Gemini models, and Antigravity's for Gemini, Claude and GPT-OSS. These are Google's logins for its own tools; another program using them may break Google's terms and get the account limited or suspended, so the sign-in states the risk first and goes on only if you agree. The official routes need none of this: a Gemini API key (provider google) and Vertex AI (provider google-vertex, signed in with gcloud).

Option	Default	Meaning
`geminiCli`	`true`	Gemini CLI sign-in. The Gemini models, on the free tier for a personal account or on a Code Assist licence.
`antigravity`	`true`	Antigravity sign-in. Gemini, Claude and GPT-OSS; once signed in, the list follows what the account can use.

Message preflight preflight

Mid-run messages interjection

Skill disclosure skills

Capability catalog catalog

Task frame and to-do list frame

Lessons memory

Risky command guard guard

Permission modes permissions

Hard constraint gate constraints

Injection screening injection

Tool output admission admission

Forgetting stale results forgetting

Summary-free compaction compaction

Checkpoints and rewind checkpoint

Drift monitor monitor

Language server diagnostics lsp

Mid-stream correction (experiment) ttsr

Goal mode goal

Plain-language board board

Completion check completion

Carry on after stopping short continuation

Notification routing notify

Cache warming warming

Judge connection warm-up warmup

Sub-agents swarm

Hive hive

File location locate

Bulk judging tool judgeItems

Built-in browser browser

Background commands background

Web reading and search web

Welcome screen welcome

Inherit existing setup inherit

MCP servers mcp

Capability packs packs

Sign in with Google (experimental) googleLogin

If mu is useful to you, star it on GitHub

Message preflight `preflight`

Mid-run messages `interjection`

Skill disclosure `skills`

Capability catalog `catalog`

Task frame and to-do list `frame`

Lessons `memory`

Risky command guard `guard`

Permission modes `permissions`

Hard constraint gate `constraints`

Injection screening `injection`

Tool output admission `admission`

Forgetting stale results `forgetting`

Summary-free compaction `compaction`

Checkpoints and rewind `checkpoint`

Drift monitor `monitor`

Language server diagnostics `lsp`

Mid-stream correction (experiment) `ttsr`

Goal mode `goal`

Plain-language board `board`

Completion check `completion`

Carry on after stopping short `continuation`

Notification routing `notify`

Cache warming `warming`

Judge connection warm-up `warmup`

Sub-agents `swarm`

Hive `hive`

File location `locate`

Bulk judging tool `judgeItems`

Built-in browser `browser`

Background commands `background`

Web reading and search `web`

Welcome screen `welcome`

Inherit existing setup `inherit`

MCP servers `mcp`

Capability packs `packs`

Sign in with Google (experimental) `googleLogin`