Navig

Make a specific open tab the active target (by tab index or url substring). Sticks on the session so later cdp_* calls act on it.

List EVERY open page in the browser (full inventory from raw CDP).

List live CDP endpoints on localhost (Chrome/Edge/Brave + Electron apps).

Type text into the currently focused element.

Probe the health of a specific connector and return latency + ok status. Use connector_list to find a valid connector_id — a health result says whether the connection works, not whether it has data.

List all registered connectors with their id, display name, domain, status, and capability flags (can_search / can_fetch / can_act). Returns one row per connector; use connector_health to probe one, or call its connector_<id>_search tool to read its data.

List durable coordination messages waiting for a participant. Read-only; returns each message id, sender, recipient, delivery state and body. recipient is a participant ref (session:<id> | agent:<id> | mission:<id> | task:<id> | formation:<id> | operator). Delivery states: persisted, delivered, consumed, acknowledged, acted, cancelled, failed, expired.

Execute an AutoHotkey v2 script via AutoHotkey.exe. Requires desktop_permission in active mission step.

Launch, resize, or bring a Windows application to the foreground. Requires desktop_permission in the active mission step.

Click a Windows UI element by its native window handle. Requires desktop_permission in the active mission step metadata.

Read the current Windows clipboard contents as text.

Write text to the Windows clipboard.

Manage file system operations with eight modes: 'read' (read text file contents with optional 1-based line offset/limit), 'write' (create or overwrite a file; set append=true to append), 'copy' (copy file or directory to destination), 'move' (move or rename file/directory), 'delete' (delete file or directory; set recursive=true for non-empty dirs), 'list' (list directory contents with optional glob pattern filter), 'search' (find files matching a glob pattern under a root directory), 'info' (get file/directory metadata: size, dates, type). Use absolute paths. Relative paths are NOT supported.

Search the Windows UI Automation element tree for elements matching name, class_name, and/or control_type.

Move the mouse to coordinates, with optional drag from a start position. Requires desktop_permission in the active mission step.

Enter text into multiple input fields. Provide a list of [x, y, text] triples. Requires desktop_permission in the active mission step.

Click multiple screen coordinates, holding Ctrl between clicks to multi-select items (files, checkboxes, list entries). Requires desktop_permission in the active mission step.

Send a Windows toast notification with a title and message body.

Execute a PowerShell command or script on the local machine and return its output. stdout and stderr are captured and returned together with the exit code.

Terminate a running Windows process by PID *or* name. When ``force`` is true, SIGKILL is used instead of SIGTERM.

List running Windows processes. Optionally filter by name substring. Returns pid, name, cpu_percent, memory_mb, and status for each match.

Delete a Windows Registry value or key. If ``name`` is omitted, the entire key is deleted.

Read a single Windows Registry value. Path format: ``HKCU:\\Software\\MyApp`` or ``HKLM:\\SOFTWARE\\...``.

List sub-keys and values under a Windows Registry key path.

Create or update a Windows Registry value.

Fast screenshot-only desktop snapshot. Skips UI element extraction. Returns cursor position, window list, and a PNG image. Use desktop_snapshot when you need interactive element coordinates. Requires desktop_permission in the active mission step.

Scroll at the specified screen coordinates. Requires desktop_permission in the active mission step.

Set the value of a Windows UI element by handle. Requires desktop_permission in active mission step.

Send a keyboard shortcut (e.g. 'Ctrl+C', 'Alt+F4', 'Win+D'). Uses AutoHotkey Send syntax. Requires desktop_permission in the active mission step.

Capture the current desktop state: screenshot, cursor position, open windows, and interactive UI elements. Set use_vision=true to include a screenshot. Set use_ui_tree=false for a faster vision-only pass. Requires desktop_permission in the active mission step.

Dump the Windows UI element tree to the specified depth.

Type text at the current cursor position (or optionally click coordinates first). Supports clearing and pressing Enter. Requires desktop_permission in the active mission step.

Act in a tab with trusted input — up to 12 steps in ONE call. Returns per-step results and (observe=true, the default) the page outline afterwards, so you rarely need a separate dock_snapshot. Pass doc from the snapshot you read so a page that changed under you is refused instead of clicked blindly.

Console messages and uncaught errors of a tab. The first call starts capturing (reload the page with dock_act navigate reload to see load-time errors).

Requests a tab made (url with secrets redacted, method, type, status, failures). The first call starts capturing.

Open a new tab in the user's browser inside the 'NAVIG agents' group and return its tab_id.

Hand a tab back when you are done. Tabs you opened are closed (close=false keeps them for the user); tabs the user shared are only removed from the agents group, never closed.

A JPEG of the tab's visible area (or one element: ref), at most max_width px wide. Works in a background tab. Prefer dock_snapshot for acting — use this when layout or images matter.

Read a tab as a compact accessibility outline: one line per element with [ref_N] handles for dock_act. Passwords and card numbers are redacted. Returns url, title, doc, rev and tree — or, with since=<rev you hold>, only a diff (added / changed / removed lines).

Is the user's own Chrome (NAVIG Dock) connected for agent control? Shows paused/connected state and latency. Use the dock_* tools to act in the user's real, logged-in browser; use cdp_* only for a separate debug browser.

List the tabs agents may use: the 'NAVIG agents' tab group in the user's browser (tabs you opened plus any the user dragged into the group). held_by says which agent is working in a tab ('you', another agent, or null = free); acting in a tab another agent holds fails as busy until it releases it or 10 idle minutes pass.

The readable text of a tab's main content (article/main or body).

Crawl a site using Firecrawl and return crawl job metadata.

Scrape or crawl a URL using Firecrawl. Returns markdown-oriented content and metadata.

Search the web using Firecrawl with optional inline scrape.

Soft-delete a stored key fact by its ID.

Store a key fact in persistent memory.

Search persistent memory for key facts matching a query.

Return counts of total, active, deleted, and superseded facts.

Update the content of an existing key fact by its ID.

Download audio locally (yt-dlp + FFmpeg) from the current track, a query, a playlist, the queue, liked songs or an explicit id/URL. Files are tagged, verified with ffprobe and provenance-stamped; they are personal offline copies and music_upload refuses to re-upload them. Use it when the operator wants the audio offline; prefer music_import to match a local folder against the catalog instead of re-downloading it. Returns the written paths, sizes, sha256, tags and a manifest.

Export observed YouTube Music state/history as JSON, JSONL, CSV, Markdown or a checksummed archive. Returns the written path(s), row counts and sha256; credentials and cookies are never included, and a history section states whether it is complete. Use music_now_playing when only the current track is wanted instead of a full export.

Inspect local audio files (tags, duration, codec) and match them against the YouTube Music catalog. Read-only: it never plays, likes or uploads. Returns per-file matches with confidence and unresolved files, so the operator can decide what to do with music_play or music_upload.

Save or remove an album/playlist from the library (verified by the header control's aria-pressed state), list the operator's uploads, or delete an upload. Use music_playlists for playlist edits and music_play for playback; artists use Subscribe instead of Save, and delete_upload is destructive (yes=true). Returns the observed before/after state.

Skip to the next track in YouTube Music and verify the track id changed. Returns the new now-playing track and the before/after ids; use music_queue to inspect what is coming next.

Return the current track and playback state from YouTube Music as typed JSON (id, title, artists, album, position, liked). Returns a null track with a warning when nothing is active; use music_play to start something or music_search to find it.

Pause YouTube Music playback and verify the app reports paused. Returns playing=false and the track id; use music_next or music_play to change what is playing.

Play music in YouTube Music and verify the resulting track. Resolves a query through the app's search; an ambiguous query fails with AMBIGUOUS_QUERY and candidates rather than silently substituting another track. Returns the now-playing track and verification evidence; use music_now_playing when you need to confirm the state later instead of assuming this call stuck.

Read or control the app's player extras: speed (read/set, verified from the player rate), skip (forward 30 s / back 10 s, progress-verified), radio/mix from the current track or a page row (music_play starts a specific track instead), autoplay (read/set, verified), lyrics, comments, stats (local observed events only) and stats_for_nerds. Returns the observed value/effect; comments and stats never claim completeness.

List the operator's YouTube Music library playlists, or show one playlist's tracks. Returns typed playlist rows with ids; use music_play with type=playlist when the operator wants one to start playing instead of only being listed.

Inspect or extend the YouTube Music queue. Returns queue items with ids; action=add resolves a query through music_search and refuses duplicates (idempotent), so retrying cannot double-add. Use music_play to start a track instead of queueing it.

Search the YouTube Music app's own interface and return typed candidates with ids (songs, albums, artists, playlists). Returns the candidate list; pass the chosen query to music_play with type=… or use music_play's pick.

Check the operator's YouTube Music desktop app: installed/running, verified transports, API pairing and what is playing. Returns the status view; pair it with music_now_playing for the track record and music_search when a capability is missing.

Upload audio files the operator OWNS to their YouTube Music library through the app's own Upload music chooser. Requires owned=true (rights attestation) and yes=true (write confirmation); refuses files NAVIG downloaded from streaming (music_download marks that provenance). Uploads are processed asynchronously: status=pending means the files were handed over but the library has not shown them yet — do not claim success. Returns the injected files, sha256, policy and verification.

Choose music from the operator's words plus local time/occasion and observed listening behaviour: 'make me happier', 'cyberpunk focus', 'christmas', tired at 23:00. Every choice is explainable (action=explain shows the signals/why) and learns from feedback (action=feedback like/skip); action=forget erases the local store. Sessions group a vibe with a duration and fold observed behaviour back in on session_end. Returns the selection and why it was chosen (or the session/playback result); use music_play to play one specific track instead; no microphone/camera/biometrics are used.

Queue a remediation restart action for an agent component.

Create a new autonomous agent goal.

Cancel an autonomous agent goal by ID.

List autonomous agent goals with state/progress summary.

Start execution for a pending/blocked autonomous agent goal.

Analyze recent agent debug/remediation logs and return recurring error patterns with recommendations.

List persisted remediation actions and recent remediation log entries.

Retry a remediation action by ID.

Install NAVIG agent as an OS service.

Get OS-level service status for NAVIG agent (systemd/launchd/windows service).

Uninstall NAVIG agent OS service.

Get autonomous agent runtime/config status including mode, personality, workspace, PID, and running state.

Get detailed information about a specific application including config, paths, and status.

Acquire a Bay item by slug (call navig_bay_list first for the slug). Installs github-backed items (space/skill/formation/prompt/webapp) and plugins; a persona returns action 'activate', a block returns 'apply' (blocks execute with inputs/approvals — never auto-run), a lens returns 'manual'. Paid/locked items return action 'locked' with the Harbor tier required — a locked item is never installed.

Browse the NAVIG Harbor Bay marketplace — installable spaces, skills, personas, blocks, plugins, and more. Returns each item's slug, name, kind, one-line tagline, whether it's unlocked for the current license, and whether it's already installed. Filter with 'kind' (space/skill/persona/block/plugin/…) or 'surface' (cli/deck/os/…).

Apply a block — run its outcome end-to-end and return a verified receipt. Secrets resolve from env/vault only. Destructive steps require their step ids in 'approvals'. Use dry_run to preview the plan without executing.

List installed NAVIG blocks (installable, verifiable outcomes you apply). Returns id, version, category, verify kind, whether paid, and the input schema.

Show one block's full spec: typed inputs, ordered steps with computed risk, and the verify criterion.

Show the usage of one installed NAVIG command: arguments, options, whether it accepts --json and its approval level. Returns the usage line and option flags; the path must be exact, as returned by navig_capabilities_find.

Find the installed NAVIG commands that match a goal in plain words, e.g. 'upload a clip to youtube' or 'remove an image background'. Read-only: it only searches the installed command list. Returns command paths with summary, owning plugin and approval level; pass one to navig_capabilities_command_help, then use navig_run_cli to execute it.

List the installed NAVIG plugins with their CLI verbs and command counts, and the accounts, keys and devices each can use now (social networks, mailboxes, media provider keys, phones). Presence only, never a secret. Returns plugins with accounts and states; use navig_capabilities_find for exact commands.

Suggest next steps for a file the agent produced or received (image, video, audio, fan-out brief), limited to installed commands and connected accounts: edit, re-score audio, review on Telegram, post per account. Returns ready steps with arguments and the steps a missing connection would unlock; run one with navig_run.

Get full NAVIG context including recent errors, active hosts, and system state for AI debugging.

Get detailed information about a specific SSH host including connection settings, paths, and apps.

List all configured applications in NAVIG. Returns app names, types, and associated hosts.

List configured database connections in NAVIG.

List all configured SSH hosts in NAVIG. Returns host names, addresses, and connection details.

List all pages in the NAVIG wiki knowledge base.

Read the content of a specific wiki page.

Run one installed NAVIG CLI command with a list of arguments and return its exit code, stdout and stderr (JSON parsed when the command printed JSON). The path must be an exact installed command from navig_capabilities_find; no shell is used. Publishing, sending and deleting commands require approval.

Execute a NAVIG CLI command and return the result. Use for any navig operation.

Create a new Mission.

List Missions, optionally filtered by node or status.

List registered NAVIG Nodes.

List ExecutionReceipts, optionally filtered by node or mission.

Advance a Mission state. Actions: start, succeed, fail:<msg>, cancel:<reason>, timeout.

Compute TrustScore for a Node from its receipt history.

Search NAVIG's local documentation for commands, guides, and configuration help.

Search the NAVIG wiki knowledge base for relevant documentation, guides, and notes.

What an agent may read from this Space's relations: provenance-carrying file items (space, rel, path, mtime), unavailable edges, character budget and dedupe already applied. Same engine as `navig space context --json`.

The active (or given) Space: id, path, phase title and relation counts. Provenance anchor for every other space tool.

Read ONE file from a Space through the linked-context guard: the target must be the active Space or reachable through its parent/link edges, the path must match that edge's include/exclude globs, must not hit the deny list (vault/env/credentials/sessions/state/databases/logs) and must resolve inside the Space's own folder (junctions cannot escape). Denied reads answer with a reason, never content.

A Space's logical graph — parent, links, children, inbound links, availability (unavailable targets carry a reason, never an error) and a parent_cycle flag. Same payload as `navig space related --json`.

Fetch a URL and extract readable content. Converts HTML to markdown or plain text for analysis.

Search the web for information. Uses Firecrawl first (if configured) with runtime fallback to other providers.

Delegate intellectual work to the Perplexity desktop app and return the finished result as structured content (text + citation URLs + thread identity). Research, investigation, analysis, comparisons, critique, translation, brainstorming. The model and reasoning mode are routed from the task; the conversation is reused for the same project/topic and can be continued later with `thread`.

Reports what this Perplexity build and account actually expose — modes (search, deep research, learn step by step, model council, computer), file upload, connectors, image and video generation — one row per capability with the UI evidence it came from and a status of verified / gated / unsupported / unverified. Returns the cached record unless refresh is true. Call it before promising a multimodal or connector workflow, or when perplexity_ask rejects a mode or connector; use perplexity_models for the model picker and perplexity_ask to delegate the work itself instead of guessing at capabilitie…

Read or set the app's model explicitly (bypassing routing). Omit name to read it.

List the models this account can select, with the routing traits (thinking, writing, reasoning, research, speed) and the gated tiers.

Check whether the app is attachable, which conversation is open and which model is set.

The persistent conversation registry: list delegated conversations, show one, resume it in the app, archive it, or clean up. A conversation survives an agent restart — continue it with perplexity_ask thread=<alias>. Cleanup: action=prune deletes this session's conversations from the app's history too (confirm=true), action=sessions shows which sessions left conversations behind.

Write publication-shaped content with the Perplexity app: an article, launch or marketing copy, a social post, a report, a translation, a rewrite. Same engine and result shape as perplexity_ask, with an editorial default task.