An AI butler tailored for personal computers - paying tribute to Jarvis from "Iron Man". It engages in conversation with you, assists you in operating your computer, resides in the background ready to respond to your summon, and is capable of both listening and speaking.
| Files | Last commit | Last update |
|---|---|---|
| 8 days ago | ||
| 19 hours ago | ||
| 1 day ago | ||
| 1 day ago | ||
| 1 day ago | ||
| 1 day ago | ||
| 1 month ago | ||
| 3 months ago | ||
| 1 day ago | ||
| 3 months ago | ||
| 3 months ago | ||
| 3 months ago | ||
| 1 day ago | ||
| 14 days ago | ||
| 20 days ago | ||
| 1 hour ago | ||
| 3 months ago | ||
| 1 month ago | ||
| 24 days ago | ||
| 3 months ago | ||
| 15 days ago | ||
| 19 hours ago | ||
| 1 hour ago | ||
| 1 month ago | ||
| 3 months ago | ||
| 19 hours ago | ||
| 19 hours ago | ||
| 19 hours ago | ||
| 1 day ago | ||
| 8 days ago | ||
| 1 month ago | ||
| 1 month ago | ||
| 1 day ago | ||
| 1 day ago |
J.A.R.V.I.S.
🌐 简体中文 | English
Just A Rather Very Intelligent System
"At your service, sir."
J.A.R.V.I.S. is a proactive AI butler that lives on your personal computer — a tribute to JARVIS from Iron Man. It resides in your operating system and can see, listen, speak and act: conversation is just the entry point, getting things done is the goal. Say "submit the project report by next Friday" and it registers the deadline with tiered reminders; say "stand down" and it waits quietly, greeting you the next morning with a daily briefing: "Good morning, sir."
North star: give everyone a computer butler that truly knows you, proactively takes care of things, and is always at your service. Not a chat box, not a voice assistant — it gets things done before you even ask.
Why JARVIS?
| The pain today | JARVIS's answer |
|---|---|
| Many AI chat tools, but few can actually operate your computer | 70+ built-in tools (100+ once MCP / harness servers are added): files, commands, browser, mouse/keyboard control, screenshot reading, camera, email — a ReAct loop plans and executes autonomously, GUI automation clicks and drags like a human |
| Cloud assistants don't know your machine, and your data isn't yours | Local-first: it lives in your OS, reads your files, drives your apps; memory, logs and configs stay in local ~/.jarvis, keys go into the OS credential manager |
| Most assistants forget you every session | Three-layer memory: session persistence + long-term memory + profile distillation (auto-learns your habits) — it knows you better the longer you use it |
Showcase
jarvis-desktop workbench — three theme skins · three columns (left: console / center: chat / right: task center · usage · system status)
![]() Fluorescent Green · task center & system status |
![]() Electric Blue · settings panel |
![]() Metal Silver · settings panel |
Terminal REPL — a conversation about memory: "Jarvis, how much do you know about me?" (skill loading → MCP knowledge graph → profile memory)
![]() Boot: Arc Reactor + MCP/LSP ready |
![]() Thinking: planning multi-source lookup |
![]() LoadSkill: loading a skill pack |
![]() ToolSearch: discovering MCP tools |
![]() Querying the knowledge-graph profile |
![]() Thinking: consolidating profile data |
![]() Answer: three-layer memory, knows you better |
![]() /models management (11 vendors) |
![]() /tts-voice management |
Voice — /voice turn-by-turn chat (Space to barge in, auto-standby after 10s silence) · /talk realtime duplex (speak to interrupt)
![]() /voice: listen → check time → gentle reminder → barge-in |
![]() /talk: startup banner (server_vad + AEC) |
![]() /talk: telling a programmer joke |
Cross-device — scan a QR code to take over the same session (/connect-phone)
![]() Terminal QR + two-way message sync |
![]() Phone: visible thinking process |
![]() Phone: MCP tool calling for weather |
WeChat ClawBot — boss your butler from WeChat anytime (/connect-wechat)
![]() Terminal QR pairing (24h valid) |
![]() Chat right inside WeChat |
![]() Full tool capabilities on the go |
Core Capabilities
| Capability | In one sentence | Docs |
|---|---|---|
| 🖥️ Terminal-native Agent | Rich terminal REPL + 70+ tools (100+ with MCP/harness) + ReAct orchestration, Ctrl+C interrupts at any stage | Concepts · Commands |
| 🎙️ Listens & speaks | /voice turn-by-turn chat (Space barge-in, auto-standby after 10s silence), /talk full-duplex realtime (speak to interrupt), TTS announces proactive reminders |
Voice |
| 🖱️ Operates your PC | Mouse/keyboard/screenshot/window tools + visual template locating, drives any software like a human | GUI Automation |
| 🗓️ Proactive butler | Daily briefing, timed reminders, deadline tracking, system monitoring — it acts without being asked | Proactive System |
| 👀 Working-context perception | Seven event sources (foreground window / folder pile-up / repo state / important mail / camera presence / upcoming calendar / system resources) feed one event bus; the notify budget decides when it is actually worth speaking up (every source off by default, you are told once after connecting if you never chose, and everything collected is exportable for self-audit) | Percept sources |
| 🛎️ Butler services | jarvis-desktop title-bar entry: email service (view/edit the current mailbox + configure SMTP + send mail) calls the backend tools directly (more to come) | Desktop & Remote |
| 🧰 Extension services | jarvis-desktop title-bar entry: a right-column panel managing MCP servers (add/edit/remove + enable/disable), built-in tools, skills, plugins and harness tools (marketplace search/install/uninstall/toggle) in one place | Desktop & Remote |
| 🧑🤝🧑 Agent team | jarvis-desktop right-column block: complex tasks are split across subagents/teammates running in parallel, each agent shows its status and current action live, can be stopped individually and messaged when it is a teammate; same implementation as the terminal /tasks /agents tools | Multi-agent collaboration |
| 🧭 Plan before acting | The model can enter plan mode on its own (read-only survey → submit a plan → restore write access); the plan is saved to .jarvis/PLAN.md and the desktop shows its summary and preview as a system message. You can also pick “规划” manually in the composer |
Desktop guide · Plan mode (中文) |
| 🧠 Three-layer memory | Session persistence (saved after every turn; /sessions and /load show created / updated time) + long-term memory + profile distillation, knows you better over time; long-term memory is visible and editable in desktop Settings → Long-term Memory, and you can just say “remember …” |
Memory |
| 🛰️ Task cockpit + web crawling | A pinned panel shows up only while it is actually doing something (chit-chat never pops it; ✕ just hides it) — at a glance: which step it is on, what failed, how long it took, and clicking any step opens the details; every running tool and every subagent gets its own small card. On the far left a second panel covers web work: pages it fetched listed by depth, extracted fields laid out as a table, and how pages link to each other drawn as a graph you can pop out and enlarge. Crawl walks the site itself (page/depth/same-domain limits, honours robots.txt) and switches to a real browser when a page is just a JS shell; anti-bot challenges are reported as-is, never bypassed. All of it can be turned off with the [visualization] switch |
Cockpit plan · Multi-task cards · Agent cards (中文) |
| 🖼️ Image cockpit | Screenshots, camera shots and page captures also land in their own window (top-left by default, draggable, position remembered): the original image large, a thumbnail strip for history, and metadata with source, size and path. The terminal could only send images to the model — you never saw them; this closes the "for humans" half, so when the model says "I looked at the screenshot" you can check it saw the same one. It does not change what is sent to the model, nor the model-side receipt wording (attached ≠ seen) | Image cockpit plan (中文) |
| 📱 One session, many ends | jarvis-desktop shell / phone QR / WeChat ClawBot share the same session | Desktop & Remote |
| 🧭 Butler runtime layer | Path-scoped no-confirmation rules in permissions.yaml + authorization leases (stop interrupting me) + notify budget (cut the noise) + /away unattended queue for approval |
W1 runtime layer (中文) |
| 🧾 The layer that makes it safe to let go | Outbound mail waits in a 30-second send window (/undo cancel stops it, and equivalent outbound commands like curl stop being auto-allowed while a hold is open), files it wrote for you roll back per turn with /undo; long-lived goals survive across sessions via the task spine (zero LLM calls while it waits), and one sentence is enough to create one — /tasks add … or just ask J.A.R.V.I.S |
W3 execution layer (中文) |
| 🧩 Open ecosystem | MCP, plugin marketplace, Skill packs, CLI-Anything (turn any software into a tool), multi-agent collaboration | Ecosystem |
| ⚙️ Config without editing files | All 207 config items are askable: don't know what one does or where to change it — just say "the CPU threshold is too high"; it looks it up and tells you what the setting does, its current value, and whether it should change it or you should (62 it can change on your behalf — 29 directly, 33 only after confirming first; the remaining 144 it just names the file and section for you). /config explain <keyword> for the same self-check |
Config system (中文) |
| 🩺 Knows what it is missing | /health caps probes 13 capabilities for real (packages / binaries / browser kernel / credentials; --deep actually launches a browser and captures one screenshot as evidence). Tools whose prerequisites are missing are blocked before the call, with what is missing, the measured evidence and how to fix it — and the receipt states "this call did not run; do not report it as verified". /repair lets it reinstall what is missing (four whitelisted actions + six gates: over 200MB asks even in yolo, /away always escalates to pending, a fix only counts as done after a re-probe). When the model has no vision the prompt switches to fallback verification routes, and screenshot receipts say "attached", never "seen" |
Capability self-check (中文) |
| 🔐 Security baseline | Five-layer permissions + four-level sandbox + operation audit + keyring encryption, hard-blocks dangerous ops | Concepts |
Provider freedom: 11 vendors supported (Alibaba DashScope / DeepSeek / OpenAI / Zhipu / Anthropic / Kimi / MiniMax / SiliconFlow / Xiaomi MiMo / Google Gemini / custom-compatible services); double-click to edit in /models, hot-switch while running.
📖 The detailed guides under
docs/guide/are currently written in Simplified Chinese; the architecture docs and code comments follow the same language.
Quick Start
# Install (pick one, Python 3.11+)
pip install "jarvis-agent[all]" # PyPI, full features
uv tool install "jarvis-agent[all]" # uv (faster)
npm install -g @acefelix/jarvis # npm (needs Node 18+ / Python 3.11+)
jarvis --init # Interactive setup: pick vendor → confirm model → enter key → auto-test → save
jarvis # Chat away: natural language directly, "/" for command palette, Tab to complete
- Defaults to Alibaba DashScope (
qwen3.7-plus, multimodal); or justexport DASHSCOPE_API_KEY=sk-xxxand launch - No Python, just a GUI? Download the jarvis-desktop installer (Windows — ships the frozen backend, no Python needed)
- All config lives in
~/.jarvis/(settings.toml + models.toml);/modeswitches permission mode,/thinktunes deep thinking - Something not working?
jarvis --doctordiagnoses dependencies and config in one shot. Want to know what it can actually do right now? Run/health capsin a session (--deepmakes it prove a claim by really launching the browser), and/repairlets it install the missing kernel itself - Optional capabilities come in 8 extras groups (
guibrowsermcpcameravisionvoicedaemonwechat):pip install "jarvis-agent[browser,gui,voice]". The playwright browser kernel is not a pip dependency — after installing the package you still needpython -m playwright install chromium, or let/health caps browser.selfcheck --repairrun it for you - Checks for a new release in the background on startup; when found, run
/updateto upgrade (disable with--no-update-check) - Mirrored on GitCode and GitHub — every commit is pushed to both
Install paths, optional dependency groups, per-OS system dependencies and the
--servebackend entry: see the Installation Guide (Chinese). To grasp where this project is heading in one read: the Blueprint.
Platform Support
Legend: ✅ verified on that OS by real-machine testing | 🟡 code-adapted and works once deps are present, but not fully tested on that OS | ⚠️ partially available (limits noted) | ⛔ not supported / not provided yet
| Capability | Windows | macOS | Linux |
|---|---|---|---|
| Chat + ReAct tool orchestration + three-layer memory | ✅ | 🟡 | 🟡 |
| File / command / Web / email built-in tools | ✅ | 🟡 | 🟡 |
| Browser automation (Playwright) | ✅ | 🟡 | 🟡 |
| GUI automation (mouse / keyboard / screenshot / window) | ✅ | 🟡 | ⚠️ window mgmt limited (pygetwindow) |
| Camera capture / gesture vision (mediapipe) | ✅ | 🟡 | 🟡 |
Realtime voice /voice · /talk (STT+TTS+AEC) |
✅ | 🟡 needs portaudio | 🟡 needs portaudio19-dev (all three need a DashScope key) |
| Proactive butler (briefing / reminders / deadlines / system monitor) | ✅ | 🟡 | 🟡 |
| W2 perception event bus | ✅ | 🟡 | 🟡 |
| Foreground-window source | ✅ | 🟡 no process name | ⚠️ depends on pygetwindow, limited |
| Calendar source (upcoming events) | ✅ Outlook COM / ICS | 🟡 ICS only (no Outlook COM) | 🟡 ICS only |
| Sandbox execution | ✅ Job Object | 🟡 sandbox-exec | 🟡 rlimits |
| MCP / Skill / Plugin / harness / multi-agent / serve backend | ✅ | 🟡 | 🟡 |
| One session, many ends (phone PWA / WeChat ClawBot) | ✅ | 🟡 | 🟡 |
Capability self-check /health caps · self-repair /repair |
✅ | 🟡 | 🟡 |
System maintenance /disk /clean /health |
✅ | ⛔ Windows-only (PowerShell + SecurityCenter2 + drive letters + cleanmgr) | ⛔ Windows-only |
| jarvis-desktop installer | ✅ NSIS .exe |
⛔ not provided yet | ⛔ not provided yet |
⚠️ All development and real-machine verification happen on Windows; macOS / Linux are code-level adaptations without full real-machine testing — Windows is recommended first. System maintenance (clean / AV / health) and the desktop installer are Windows-only today; per-row limits are flagged above. Dependency details: Installation Guide (Chinese). “Windows-only, honestly rejected on POSIX” is not just prose: CI (Linux) reconciles the built-in cleaner rulebook entry by entry, proving it rejects exactly those Windows-semantics paths while every other validation gate stays in force (
tests/test_release_contract.py).
Documentation Map
| Topic | Doc |
|---|---|
| Start here: the butler's own onboarding guide (persona-style) | USER_GUIDE.md (中文) |
| Install / upgrade / platform deps | docs/guide/installation.md (中文) |
| All commands / model management / deep thinking | docs/guide/commands.md (中文) |
| Voice chat / realtime duplex / voices | docs/guide/voice.md (中文) |
| Desktop shell / proactive reminders / phone / WeChat / serve protocol | docs/guide/desktop.md (中文) |
| Permissions / compaction / memory / sandbox / self-healing | docs/guide/concepts.md (中文) |
| Authorization leases / notify budget / unattended mode | docs/architecture/16-W1判断层.md (中文) |
Reversible actions (/undo) / cross-session task spine |
docs/architecture/18-W3兑现层.md (中文) |
| Multi-agent / plugins / CLI-Anything / email | docs/guide/ecosystem.md (中文) |
| Config reference / vendor setup / FAQ | config-docs/ (中文) |
Capability self-check / missing-dependency gating / self-repair (/health caps, /repair) |
docs/architecture/20-能力自检与自我修复.md (中文) |
| Architecture (20 chapters) | docs/architecture/ (中文) |
| Testing / CI / publishing / directory layout | docs/guide/development.md (中文) |
| Blueprint (vision · capability model · roadmap) | docs/BLUEPRINT.md |
Development Roadmap
Stages 8–11 (policy / perception / execution / experience) and the full capability list: Blueprint; execution-level detail: docs/roadmap/.
Contributing
Issues and PRs are welcome — docs, tests and translations count too.
- Read CONTRIBUTING.md first (dev setup, test commands, code conventions, PR rules)
- Bug reports: use the issue template with
jarvis --version, platform, repro steps and redacted logs - Feature requests: open an issue to align on direction first
- Code changes: fork → branch → commit → PR; CI (syntax gate + full pytest + gitleaks + Markdown link check) must pass
- Security issues: do not open a public issue; follow SECURITY.md
By participating you agree to the CODE_OF_CONDUCT.md.
Disclaimer
J.A.R.V.I.S. is still in development and validation; features are not yet fully stable. If you hit problems or rough edges, feedback is welcome at 13985465782@163.com. Every piece of feedback drives improvement — thanks for your support and tolerance!
Thank You for Your Support
Thank you for using J.A.R.V.I.S.!
"J.A.R.V.I.S. — At your service, sir."
| X (Twitter) | Douyin | Bilibili | |
![]() |
![]() |
![]() |
![]() |
Jarvis Agent App Development Community
Scan to join and exchange Agent app development experience, plugins and various AI use cases with other Agent app peers.
Introduction
An AI butler tailored for personal computers - paying tribute to Jarvis from "Iron Man". It engages in conversation with you, assists you in operating your computer, resides in the background ready to respond to your summon, and is capable of both listening and speaking.
Customize your domain























