convez.ai
MCP Server
Dhanush-Aries

jarvis

Dhanush-Aries/jarvis

Ver no GitHub
InicialPluginScore · de 0 a 100

Portable multi-provider AI assistant — voice, web, CLI, daemon — works with zero API keys via Claude Max/Ollama; MCP-powered coding, security, and desktop automation agents

Instalação

$npx convez-ai Dhanush-Aries/jarvis

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
2
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
v1.0.0
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Jarvis

A portable, multi-provider, multi-interface autonomous AI assistant — your "Avengers Jarvis". One core kernel; four ways to talk to it (terminal, web, voice, daemon); every LLM provider behind a single router; works on any environment — including with zero API keys, running on a Claude Max/Pro subscription (via the claude CLI) or a local Ollama fallback.

What it does

  • Runs on your Claude subscription — no API key, no per-token billing. Models prefixed claude-code/* shell out to the claude CLI and use your Claude Max/Pro OAuth login. This is the default in config.yaml.
  • Talks — optional "hey jarvis" voice loop (wake word → whisper STT → reply → neural TTS).
  • Codes — an engineering agent that reads, writes, and runs code.
  • Hacks — an authorised-pentest agent wired to your MCP security tools.
  • Drives the desktop — OpenClaw gives Jarvis hands on the PC: launch apps, open files/URLs, type, hotkeys, screenshots, clipboard, browser. Fine in-page browser control via Playwright MCP.
  • Automates anything, end to end — Hermes is an autonomous operator that chains desktop + browser + shell + code together to finish a goal.
  • Any provider — Anthropic, OpenAI, Groq, Gemini, DeepSeek, Ollama… auto- detected; missing keys are skipped; Ollama is always the last-resort fallback.

Quick start

cd ~/jarvis
./bootstrap.sh            # core only      (or: ./bootstrap.sh all)
source .venv/bin/activate
jarvis doctor            # what's available on this machine
jarvis chat              # talk to it

No API keys? If you're logged into the claude CLI (Max/Pro plan) it just works — config.yaml defaults every tier to claude-code/sonnet. No subscription either? Install Ollama (ollama pull llama3.1) and it falls back to local. Have API keys? Put any subset in .env and the router can prefer them.

Providers & the Claude Max story

The router walks each task tier top-to-bottom, skipping any model whose backend is unavailable, and always ends at a reachable Ollama.

  • claude-code/* (e.g. claude-code/sonnet, claude-code/haiku) — routes through the claude CLI using your Claude Max/Pro subscription (OAuth, no API key). Plain completions run as:
    claude -p --output-format json --strict-mcp-config --setting-sources user
    
    The last two flags are essential for speed: without them every call loads ~20 MCP servers plus the project CLAUDE.md (minutes per call instead of ~4s).
  • claude-* (no prefix, e.g. claude-sonnet-4-6) — the paid Anthropic API via LiteLLM, used only if ANTHROPIC_API_KEY is set.
  • Everything else (openai/*, groq/*, gemini/*, deepseek/*, …) — via LiteLLM when the matching key is present.
  • ollama/* — always-available local fallback.

The subscription completion path has no native function-calling. To actually act on the Max plan, the claude.code skill delegates a full agentic Claude Code run — its own tools, MCP servers, and skills — that can read/write files and run commands. It defaults to autonomous=true (full-auto). Tiers default to claude-code/sonnet; edit config.yaml to change them.

Interfaces

jarvis chat              # interactive terminal (default)
jarvis ask "fix the bug in app.py and run the tests"
jarvis ask --agent hacking "recon example.com"
jarvis ask --agent hermes "open my email in the browser and summarise the top 3"
jarvis web               # local dashboard at http://127.0.0.1:8787   (needs [web])
jarvis daemon            # scheduled/autonomous background jobs        (needs [daemon])
jarvis voice             # "hey jarvis" voice loop                     (needs [voice] + mic)
jarvis skills            # list every registered skill (incl. MCP tools)
jarvis doctor            # capability report (see below)

Voice ("hey jarvis")

jarvis voice runs a wake-word loop: it listens for the wake word (jarvis, configurable), then records your request using energy-based voice-activity detection (records until you stop talking, not a fixed timer), transcribes, answers, and speaks the reply.

  • STT — faster-whisper (model size configurable, default base.en).
  • TTS — layered backends, best first:
    1. edge-tts — Microsoft neural voices, online, no model download (preferred).
    2. piper — fully offline neural TTS (needs a downloaded .onnx voice).
    3. espeak / say — system fallback.

Inspired by the openjarvis project. Needs the [voice] extra and a microphone; if either is absent it exits cleanly with a clear message and the other interfaces are unaffected.

Agents

The router tags each request — by a free keyword fast-path first, then a cheap classifier model — and dispatches to one agent. Force one with --agent <name>.

AgentWhat it does
chatConcise general assistant.
codingEngineering mode — read, edit, run, verify code.
hackingAuthorised offensive-security mode (PTES), wired to MCP security tools.
automationErrands across connected tools (mail, calendar, notes, shell).
openclawDesktop operator — drives apps, browser, keyboard, screen, clipboard.
hermesAutonomous operator — chains desktop, browser, shell, code, and the claude.code delegate end-to-end.

Desktop control (OpenClaw) & browser

The computer skill category gives Jarvis hands on the machine, with cross-platform backends and graceful detection:

SkillPurpose
openclaw.launch_appOpen a desktop app by command/name
openclaw.openOpen a file, folder, or URL with the default handler
openclaw.browseOpen a URL in a (named or default) browser
openclaw.typeType text into the focused window
openclaw.keyPress a key/hotkey (e.g. ctrl+t, Return)
openclaw.screenshotCapture the screen to a PNG
openclaw.clickMove the mouse and click
openclaw.clipboard_set / openclaw.clipboard_getRead/write the clipboard

Backends are detected per session — Wayland: wtype / grim / wl-copy; X11: xdotool / scrot / xclip; macOS: built-ins. Each skill reports clearly if a backend is missing instead of failing silently.

On this Wayland machine, keyboard, screenshot, clipboard, app-launch, and browser launch all work. Mouse-click needs ydotool (sudo pacman -S ydotool && systemctl --user enable --now ydotool).

For fine in-browser control (navigate, click, fill, snapshot) use the Playwright MCP server, surfaced via the MCP bridge as mcp.playwright.*.

Architecture

interfaces/{cli,web,daemon,voice_app}   thin frontends — no business logic
        │  build RequestContext
        ▼
core/kernel.py   ── route ──►  agents/{chat,coding,hacking,automation,openclaw,hermes}
        │                              │ ReAct loop + autonomy gate
        │                              ▼
        │                       skills/  builtins + computer(openclaw) + claude.code
        │                                + MCP bridge + project bridge + plugins
        ▼
providers/router.py   tier-based model choice + fallback
        │   claude-code/* → claude CLI (Max/Pro)   claude-*/openai/* → LiteLLM   ollama/* → local
        ▼
memory/  SQLite: conversations, task_journal (audit), longterm (FTS5), jobs
core/capabilities.py  probes OS/keys/audio/tools/desktop → everything self-disables gracefully

Configuration

  • config.yaml — models, task tiers, autonomy mode, voice, web (non-secret).
  • .env — API keys only. Any subset; none required.
  • ~/.jarvis/ — SQLite DB + drop-in plugins/*.py (auto-loaded @skills).

jarvis doctor

Prints a capability report so you know what's live on the box, including:

  • claude_code (Max/Pro plan) — is the claude CLI present.
  • cloud_providers / ollama — which API providers and local fallback are up.
  • voice_ready — voice deps and a microphone present.
  • desktop_control — detected backends for keyboard / screenshot / mouse_click / open_app / browser (and the session type, e.g. wayland/x11/macos).

Autonomy & safety

Default mode is autonomous (chains tasks without asking). approval_required_for patterns in config.yaml are gated even so — destructive shell commands (shell.destructive) and outbound sends (net.send) are hard-stopped. The claude.code delegate defaults to autonomous (full-auto). Tool output is always treated as untrusted data, never executed as instructions. Use only against systems you're authorised to operate or test.

Docker

docker compose up --build         # jarvis (web) + ollama
docker compose exec ollama ollama pull llama3.1

Extend it

Drop a file in ~/.jarvis/plugins/:

from jarvis.skills import skill

@skill(name="weather.now", description="Get current weather for a city.",
       parameters={"type":"object","properties":{"city":{"type":"string"}},"required":["city"]})
async def weather_now(city: str) -> str:
    ...

It's available to the agents on next start.

Ver o original no GitHub

Bom para

Relacionados