CLI

Every command accepts --config PATH (-c). --auto switches Sentinel to auto mode for that run; explicit deny rules still apply.

Talking to the agent

nanomuse serve [--host 0.0.0.0] [--port 8787] [--no-qr] [--no-auth]

The always-on agent with the phone app. See app.md. --no-auth is for local development only.

nanomuse chat [--resume] [--show-thinking] [--auto]

Interactive session in the terminal with approvals inline. Slash commands:

Command	Effect
/help	list commands
/reset	clear the conversation and the taint flag
/memory, /goals	list memories / goals
/audit [n]	recent audit entries
/tools	tools the agent can use right now
/tainted	whether the session has read private data
/permissions	standing permissions you granted (key and lifetime)
/revoke <key>	take one back, e.g. /revoke shell:git

--resume continues the most recent session (<data_dir>/sessions/).

nanomuse run "Summarise the top three Hacker News stories into hn.md" [--auto]

One task, then exit. Non-zero exit if the agent fails.

nanomuse daemon [--interval 3600] [--once]

Advance every active goal, sleep, repeat. Runs in Sentinel auto mode, so pair it with deny rules and a tight egress_allowlist. The app's Background work switch does the same thing inside nanomuse serve, one goal per interval, with updates posted to the chat.

Goals

nanomuse goals list [--status active|paused|done] [--category health|finance|career|learning|…]
nanomuse goals show g_1a2b3c
nanomuse goals add "Learn Rust" -s "Read the book, ch. 1–4" -s "Build a CLI" -s "Publish a crate" \
    --category learning --due 2026-12-31 --check-in "weekly sun 19:00"
nanomuse goals run g_1a2b3c            # one background pass now
nanomuse goals status g_1a2b3c paused
nanomuse goals delete g_1a2b3c

Reminders and routines

nanomuse reminders list [--all]                                  # --all includes the recently fired
nanomuse reminders add "Call mum" --at "2026-10-01 18:00"        # one message at that time
nanomuse reminders add "Summarise unread email" --repeat "weekdays 07:30" --task   # the agent does it, then reports
nanomuse reminders cancel r_1a2b3c

A reminder says one thing at the time you named; a routine (--task) is work the agent does at that time with its tools. Both fire from the running nanomuse serve — those set from the terminal land in the main chat, those set in a side chat stay there. Cadence grammar as for goal check-ins: daily HH:MM, weekdays HH:MM, weekly <mon…sun> HH:MM, monthly <day> HH:MM.

Triggers

nanomuse triggers list [--all]                                                   # --all includes cancelled ones
nanomuse triggers add mail  "Summarise it and draft a reply" --match "landlord"   # a new mail whose sender/subject has every word
nanomuse triggers add event "Put together a one-page brief"  --match "review" --lead 30   # 30 min before a matching event
nanomuse triggers add hook  "Check that the site is up"      --match "deploy"    # prints the URL to POST to
nanomuse triggers cancel t_1a2b3c

A trigger fires when something happens rather than at a time: mail needs the email connector (the running server looks at the inbox every triggers.mail_poll_minutes, by IMAP UID, so nothing is replayed and nothing fires twice), event needs a calendar feed, hook is a URL with a key that any program can POST to — the body becomes the agent's context. Each firing is a background run in the chat the trigger was set from, shown in the Feed as New mail: …, Coming up: … or Webhook: …. The mail or the request is handed to the model as data, with the instruction that only your standing text says what to do.

Calendar

nanomuse calendar add Work "https://calendar.google.com/calendar/ical/…/basic.ics"   # link goes to the vault as CALENDAR_WORK
nanomuse calendar add Family ~/family.ics                                            # or an .ics file on disk
nanomuse calendar agenda --days 7                                                    # what is on, grouped by day
nanomuse calendar free --day tomorrow --minutes 45                                   # gaps in the working hours
nanomuse calendar feeds                                                              # each feed: events, last read, error
nanomuse calendar remove Family

Feeds are re-read every refresh_minutes by the running server; agenda --refresh fetches now. The agent has the same view through its calendar tool, plus draft, which writes an .ics the app shows as an Add to calendar card — it never writes to your calendar directly.

Contacts

nanomuse contacts search "ali"                          # by name, nickname, company, email or phone
nanomuse contacts list -n 20                            # the first people alphabetically
nanomuse contacts add "Bob Li" -e bob@example.com --note landlord   # into the agent's own book
nanomuse contacts sources                               # each address book: people, last read, error
nanomuse contacts add-source Google ~/Downloads/contacts.vcf         # a .vcf file, or a link (→ vault CONTACTS_GOOGLE)
nanomuse contacts remove-source Google

The agent has the same view through its contacts tool; doctor reports how many people it knows and from where.

Skills

nanomuse skills list                                    # built-in and yours, on or off
nanomuse skills show trip-plan                          # the SKILL.md, as the model reads it
nanomuse skills new standup-notes                       # a SKILL.md to fill in, printed with its path
nanomuse skills add ./my-skill/                         # a folder or a SKILL.md file …
nanomuse skills add https://github.com/anthropics/skills/tree/main/skills/pdf   # … or a raw link / GitHub folder page
nanomuse skills disable inbox-triage                    # out of the model's list; the folder stays
nanomuse skills enable inbox-triage
nanomuse skills remove standup-notes                    # one of yours (built-in ones are disabled, not removed)

Skills are folders with a SKILL.md in the Agent Skills format; yours live in <data_dir>/skills/ and one with the same name as a built-in replaces it. In chat, /name at the start of a message runs one; doctor lists what is loaded and any folder it could not read.

Memory

nanomuse memory list
nanomuse memory add "Prefers short answers" --category preference
nanomuse memory recall "写邮件给房东"    # what the agent would recall for this, with closeness
nanomuse memory forget m_9f8e7d        # id or a phrase to search for
nanomuse memory clear --yes
nanomuse memory tidy --dry-run         # what a tidy-up would merge and drop
nanomuse memory tidy                   # do it; every change is logged
nanomuse memory changes                # the log, newest first
nanomuse memory restore c_1a2b3c4d     # undo one change

memory recall shows the ranking the agent would get for a message: keyword hits and, when an embedding endpoint is set up (configuration → memory), hits by meaning fused in, each with its cosine closeness — the way to see whether "写邮件给房东" finds "the landlord is Bob Li" before relying on it. doctor says whether recall by meaning is on, with which model, and how many memories are indexed.

Vault

nanomuse vault set EMAIL_PASSWORD      # prompted; or --value for scripts
nanomuse vault list                    # names only, never values
nanomuse vault delete EMAIL_PASSWORD

Reference secrets as {{vault:EMAIL_PASSWORD}} in the config (connectors only; a command never receives a secret). See sentinel.md.

Phone

nanomuse phone traces [--limit 20]                 # every phone_task on record: id, when, status, steps, goal
nanomuse phone trace pt-20260923-230710-5be7       # one task step by step: screen, latency, action, what was sent
nanomuse phone trace pt-20260923-230710-5be7 -o trace.html   # a self-contained page with every screen and tap drawn on it

Traces are written whenever the agent operates the phone (gui.md); the HTML page is one file with the screenshots inlined, made to be attached to an issue.

Audit and config

nanomuse audit [-n 20] [--json]        # recent decisions, approvals, tool calls
nanomuse config init [--path config/config.toml] [--force]
nanomuse config show                   # effective settings, secrets masked
nanomuse config path                   # which file is in use
nanomuse doctor [--no-model]           # config, data dir, model, connectors — one screen
nanomuse version                       # also: nanomuse --version / -V
nanomuse mcp                           # this computer's screen and hands as an MCP server on stdio, for another host

nanomuse mcp is for a host that is not our runtime — nanoMuse on DeepSeek Harness — and serves computer_screen and computer_act with the runtime's descriptions; a step the Sentinel would ask about (Enter, a submit, a heavy shortcut, a click on a sensitive word) is refused with the reason until the call carries confirmed: true, which the model may set only after the person agreed in the conversation.

nanomuse doctor is the first thing to run when something is off, and what to paste into a bug report: which config file is in use, where the data lives, which model and endpoint are configured and whether a key is set, whether recall by meaning is on and how many memories are indexed, which web search provider answers and whether it has its key or URL, whether commands run in the sandbox (and why not, if not), the tools the agent has, connector state (mailbox, calendar feeds, address books), and a one-line call to the model with its latency (--no-model skips that). It exits non-zero when something needs fixing and says what.