Personal Jarvis
A self-described "open-source AI agent super-app for your desktop": one local app running
Claude Code, Codex, OpenCode, Kimi Code, GLM, Grok Build, Antigravity, Cursor CLI and DeepSeek
Harness in parallel terminal panes, plus persistent background agents, voice control, and computer
and browser automation. Apache-2.0 (releases ≤1.6.0 stay MIT), Python, 123★ / 52 forks, 5,659
commits on main, read 2026-10-07. No cloud account: the project runs locally and users bring their
own API keys, subscriptions, or local models.
Where it sits next to NemoClaw and Scout
The spoke already tracks a vendor runtime-governance layer above individual harnesses: nemoclaw (NVIDIA, sandboxes OpenClaw/Hermes/LangChain through OpenShell, routes inference) and Scout (Microsoft, wraps one harness with Entra identity). Personal Jarvis is a third shape of the same idea, community rather than vendor, and a desktop app rather than a CLI control plane: it runs multiple coding CLIs side by side (Claude Code, Codex and six others) in one UI, with a four-tier risk gate ("safe, monitor, ask, block") instead of NemoClaw's sandbox-plus- routed-inference model. It governs by prompting and recording, not by containment — nothing here is a sandbox per agent, and model calls go directly to the provider the user configured rather than through a managed inference layer. So the multi-harness idea now has a fourth owner (community) next to NVIDIA, Microsoft and (per kimi-claw) Moonshot's single-harness hosting play.
What's bundled, beyond the harness panes
- Persistent agents ("Jarvis Agents") with their own memory, schedules and tools; coding missions run in git worktrees and go through review before merging, the same worktree-based isolation this spoke sees in other harnesses' plan-then-execute workflows.
- Voice: custom wake word, interruptible conversation, dictation into any app, phone calls; can run fully offline on Whisper and Piper.
- Computer use: desktop and browser control under user-set permissions.
- A marketplace: 40+ plugins, MCP servers, skills and CLI integrations (GitHub, Notion, Slack, Home Assistant), the same skill-pack shape this spoke has seen in narrower, single-purpose tools (claude-seo, agentic-seo-skill) but here as a general marketplace spanning many domains at once.
- Models: OpenAI, Anthropic, Gemini/Vertex, OpenRouter, NVIDIA, Ollama, any OpenAI-compatible endpoint, plus Claude/ChatGPT subscriptions directly.
Tier and what's unverified
T3 — read from the project's own README and site; no independent review, benchmark, or security
audit found. The repo ships a desktop-ttu-baseline.json file with no description reachable from the
README, so whatever it measures is not recorded here. The macOS build is explicitly not notarized.
At 5,659 commits and 17 open issues against 123 stars, this reads as a large, actively developed
project that has not yet found a wide audience; nothing here says why.
Related
nemoclaw · openclaw · kimi-claw · agent-orchestration · agent-skills · model-context-protocol · synthesis