Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
title: "Reference: Dodaios CLI" slug: reference-jarvis-cli diataxis: reference status: active owner: maintainers last_reviewed: 2026-04-29 phase: "-" audience: developer tags: [cli, jarvis-ask, jarvis-skills, cli-tools]
Reference: Dodaios CLI
Complete reference of all command-line tools shipped with
Dodaios — the main CLI dodaios (earlier name jarvis), the voice bridge
jarvis-ask, the skill runner jarvis-skills, the review-pipeline
tools jarvis-review-eval / jarvis-review-gc, plus the catalog of the
~20 external CLI tools that the Brain-Manager can dispatch as tools
(gcloud, aws, gh, stripe, docker, kubectl etc.).
Table of Contents
- Main CLI:
dodaios - Voice bridge:
jarvis-ask - Skill runner:
jarvis-skills - Review pipeline:
jarvis-review-eval/jarvis-review-gc - CLI tool registry (external tools)
- Launcher scripts (
run.bat) - Module invocations (
python -m ...)
Main CLI: dodaios
The main entry point — installed via pip install -e . from the
repo root. Presents itself primarily as a tray/setup wrapper; the
running desktop app is started via run.bat.
Synopsis:
dodaios [-h] [--version] [--wizard] [--check] [--plugins] [--debug]
[--phase5-doctor] [--install-admin-helper]
| Flag | Effect |
|---|---|
--version |
Shows the package version and exits. |
--wizard |
Starts the setup wizard for API keys (Windows Credential Manager). |
--check |
Hardware analysis (CUDA, RAM, mic devices) + Whisper recommendation. |
--plugins |
Lists all plugin slots from the plugin registry (Wakeword, STT, TTS, Brain, Harness, Tool, Channel). |
--debug |
Console logging + verbose config dump at startup. |
--phase5-doctor |
Checks Phase-5 prerequisites: admin helper installed, vision deps available, kill hotkey bound, cost config set. |
--install-admin-helper |
Generates HMAC secret and registers the admin-helper shortcut (Phase-5 privileged-action handler). |
Example:
# Initial setup
dodaios --wizard
# Hardware check
dodaios --check
# Plugin inventory
dodaios --plugins
Source: dodaios/__main__.py:main.
Voice bridge: jarvis-ask
A thin CLI/HTTP bridge over which a Dodaios-Agent (or another CLI agent) can ask the user a question by voice. Dodaios reads the question out via TTS, captures the spoken answer, and returns it on stdout.
Synopsis:
jarvis-ask "<Frage>" # open question, free text
jarvis-ask --yes-no "<Frage>" # yes/no variant
jarvis-ask --yes-no --json "<Frage>" # JSON output for scripting
Default timeout: 30 seconds.
Exit codes:
| Code | Meaning |
|---|---|
0 |
Answer received — stdout contains the raw answer text (or full JSON with --json). |
2 |
Voice pipeline offline (e.g. JARVIS_VOICE=0 or headless mode). |
3 |
Pipeline busy (another voice turn active). |
4 |
Timeout — user did not answer within 30 s. |
5 |
Network error talking to the local backend. |
Pre-check (optional, ~50 ms):
curl -s http://127.0.0.1:47821/api/voice-bridge/health
# -> {"available": true, "state": "LISTENING"} -> jarvis-ask
# -> {"available": false, "state": null} -> AskUserQuestion-Fallback
Never query over the voice bridge:
- API keys, passwords, auth tokens (STT data leak into the audio log).
- Security-critical "Are you sure?" confirmations (yes/no ambiguous in STT).
- Payment/cost confirmations (same reason).
Source: dodaios/clis/jarvis_ask.py:main, endpoint
POST /api/voice-bridge/ask in dodaios/ui/web/voice_bridge_routes.py.
Skill runner: jarvis-skills
CLI frontend to the SkillRegistry — lists, inspects, runs skills,
imports external openclaw (Claude-Code-compatible) skills, and promotes
Dodaios-Agent-authored drafts to state=active.
Synopsis:
jarvis-skills (--list | --info NAME | --run NAME |
--import-claude-skills PATH | --list-drafts | --promote SLUG)
| Flag | Effect |
|---|---|
--list |
Table of all skills (name, version, state, triggers). |
--info NAME |
Full detail for a skill: frontmatter, triggers, tools, resources, risk tier. |
--run NAME |
Run a skill (requires an MCP connection context). |
--import-claude-skills PATH |
Imports skills from the external openclaw CLI's skill directory into the user skill repo. |
--list-drafts |
List Dodaios-Agent-authored drafts (state=draft). |
--promote SLUG |
Promote a draft skill to state=active — user-explicit activation with a security lint of the skill body. |
Examples:
# Inventory
jarvis-skills --list
# Skill detail
jarvis-skills --info memory-save
# View Dodaios-Agent drafts
jarvis-skills --list-drafts
# Promote a draft (Plan §7.5: user-explicit activation)
jarvis-skills --promote my-new-skill
Promote lifecycle (Plan §AD-8): draft → lint (security) → frontmatter
state: active → SkillRegistry reload → audit entry skill_promoted.
Throws UnsafeSkillError if the lint finds forbidden calls
(eval/exec/system etc.).
Source: dodaios/skills/cli.py.
Review pipeline: jarvis-review-eval / jarvis-review-gc
Phase-8.6 tools for the review pipeline (dispatch_with_review).
Use eval buckets from data/review/eval/.
jarvis-review-eval
Synopsis:
jarvis-review-eval [--quick] [--mock] [--bucket BUCKET]
[--report PATH]
| Flag | Effect |
|---|---|
--quick |
Subset of the eval suite (deterministic cases, ~seconds). |
--mock |
Mock worker (no API call, free, < 1 s). For CI / smoke. |
--bucket NAME |
Filter to a single bucket (code_gen_trivial, research, skill_authoring etc.). |
--report PATH |
Writes a JSON report with pass rate, latency, token usage. |
Examples:
# Mock mode (deterministic, free, < 1 s)
jarvis-review-eval --quick --mock
# Full suite (real, ~$2-5, ~10 min)
jarvis-review-eval --report data/last-eval.json
# Bucket filter
jarvis-review-eval --bucket code_gen_trivial
jarvis-review-gc
Synopsis:
jarvis-review-gc --older-than <duration>
Cleans up run artifacts under data/review/runs/<run_id>/ when they
are older than <duration>. The data/review.log JSON-Lines file
remains untouched (gapless audit trail).
Example:
jarvis-review-gc --older-than 30d
Sources: dodaios/cli/review_eval.py, dodaios/cli/review_gc.py.
CLI tool registry (external tools)
A central registry indexes ~20 external CLI tools that the Brain-Manager (main Dodaios and Dodaios-Agents) can dispatch as tools. Each entry carries a spec, auth profiles, a risk-tier default, and a probe status.
Inventory (as of 2026-04-23, dodaios/clis/catalog/seed_catalog.json):
| Name | Display | Description |
|---|---|---|
gcloud |
Google Cloud CLI | Compute, Storage, IAM, GKE, Run |
az |
Azure CLI | Resources, VMs, AKS, Functions |
aws |
AWS CLI v2 | S3, EC2, Lambda, IAM, RDS |
wrangler |
Cloudflare Wrangler | Workers, Pages, R2, D1 |
vercel |
Vercel CLI | Deployments, Projects, env variables |
netlify |
Netlify CLI | Sites, Deployments, Functions, Env |
heroku |
Heroku CLI | Apps, Dynos, Addons, Releases |
railway |
Railway CLI | Projects, Services, Deployments |
flyctl |
Fly.io CLI | Apps, Machines, Volumes, Secrets |
render |
Render CLI | Services, Deploys, Logs |
supabase |
Supabase CLI | Projects, DB, Edge Functions, Migrations |
firebase |
Firebase CLI | Projects, Hosting, Functions, Firestore |
pscale |
PlanetScale CLI | MySQL branches, deploy requests |
neonctl |
Neon CLI | Serverless Postgres, Branches |
gh |
GitHub CLI | Repos, PRs, Issues, Actions, Releases |
glab |
GitLab CLI | Repos, MRs, Issues, CI/CD, Snippets |
stripe |
Stripe CLI | Webhooks, Events, Products, Subscriptions |
twilio |
Twilio CLI | Numbers, Messages, Voice, Verify, Flex |
docker |
Docker CLI | Containers, Images, Volumes, Networks |
kubectl |
Kubernetes CLI | Cluster management: pods, deployments etc. |
gam |
Google Workspace CLI (GAM) | Calendar events, Gmail messages, Drive, Users |
Architecture components
CliCatalog(dodaios/clis/catalog/__init__.py) — loadsseed_catalog.json+ allows user-custom specs inuser_clis_dir()/custom.json.CliStatusProber(prober.py) — async probe of whether a CLI is installed and login-capable; caches status per CLI.CliAuthManager(auth.py) — manages auth profiles (OAuth CLI, ENV token, manual login); injects ENV variables per call.CliInstaller(installer.py) — winget/scoop/npm/pip installation via a manifest from the catalog.CliTool(tool.py) — one instance per connected CLI; implements the tool protocol for the Brain-Manager. One tool per CLI, not per subcommand.UsageLog(usage_log.py) — SQLite DB underuser_data_dir()/ cli_usage.db; every call is logged with latency + exit code.CliToolRegistry(registry.py) — aggregator. Set up byWebServer._setup_cli_registry()at startup; bootstrap runs asynchronously in the background.
Risk-tier defaults
From seed_catalog.json:
- payments / prod-cluster (
stripe, prod branches inkubectl,awswith--profile=prod): default tierask— user confirmation per call. - all others: default tier
monitorwith a blacklist (e.g.gcloud projects delete,aws s3 rm --recursive).
Binary guard
Every CliTool call validates before execution that the command arg
begins with the binary_name from the catalog. A brain that
attempts run-shell {"command": "rm -rf /"} would already have failed at
the risk-tier executor; the binary guard is the second
line of defense.
ENV injection
CliAuthManager.env_for(spec) returns a dict with the auth ENVs
(e.g. CLOUDSDK_AUTH_ACCESS_TOKEN for gcloud,
GH_TOKEN for gh, STRIPE_API_KEY for stripe). The
CliTool merges that into os.environ for the subprocess.
Output truncation
stdout: max 4000 characters, stderr: max 2000 characters. Prevents
token bloat with verbose CLIs (kubectl describe pod etc.). The full
output lands in the UsageLog when the user looks it up.
Launcher scripts (run.bat)
Windows launcher in the repo root. Sets the working directory, activates
.venv\Scripts\activate.bat if present, and starts the launcher
via python or pythonw.
Variants:
| Invocation | Effect |
|---|---|
run.bat |
pywebview window via pythonw (no console window), voice pipeline on. |
run.bat --debug |
Console window visible, JARVIS_DEBUG=1, verbose logging. |
run.bat --headless |
Backend-only on 127.0.0.1:47821 (no window, no voice pipeline). |
run.bat --dev |
Frontend from the Vite dev server (port 5173), JARVIS_DEV=1 — hot reload. |
Disable voice: set JARVIS_VOICE=0 before the invocation.
Module invocations (python -m ...)
Direct modules that can be invoked without an entry-point wrapper:
| Invocation | Purpose |
|---|---|
python -m dodaios |
Identical to dodaios (tray + setup). |
python -m dodaios.ui.web.launcher |
Desktop-app launcher directly; accepts --headless / --dev. |
python -m dodaios.skills.cli |
Identical to jarvis-skills. |
python -m dodaios.hardware.detection |
Standalone hardware analysis (RAM, CUDA, mic, Whisper recommendation). |
Example:
# Desktop app as backend only (for browser frontend)
python -m dodaios.ui.web.launcher --headless
# Hardware analysis as standalone
python -m dodaios.hardware.detection
Related docs
- ADR-0011: Pure Dispatcher (4 tools) — why only
run-shelland not arun-clitool. - ADR-0008: Computer-Use in-process — Computer-Use is a special case, does not run over the subprocess pattern of the CLI tools.
Sources in the code
- Entry points:
pyproject.toml [project.scripts]. dodaios/__main__.py— main CLI.dodaios/clis/— voice bridge + CLI tool registry.dodaios/skills/cli.py— skill CLI.dodaios/cli/review_eval.py,review_gc.py— review-pipeline tools.run.bat— Windows launcher.dodaios/ui/web/voice_bridge_routes.py— REST endpoint behindjarvis-ask.