Jarvis OS

claude plugin marketplace add ManceRayder42/jarvis-os
claude plugin install jarvis-os

Restart Claude Code, run /jarvis-setup, done.

The problem this solves

Claude Code already has memory. It's keyed to git root, and it silently does not load when your working directory isn't in a repo it recognizes.

There's no error. No warning. The context you expected just isn't there, and you don't find out until you're halfway through explaining something for the fourth time.

Jarvis OS sidesteps it: one hub directory, one MEMORY.md, injected at every session start regardless of where you launched from. A SessionStart hook does it, it takes milliseconds, and it exits silently on any error — a memory layer that breaks your session start is worse than no memory layer.

What you get

Core and research skills run locally and need no account and no API key. Anything that costs money ships toggled off. A free plugin shouldn't ask for a credit card in the first five minutes.

The one habit you have to build

End every working session with /done — or just tell Claude to close out the session, which triggers the same skill.

This is the only manual step, and the only change to how you already work. Everything else is automatic: memory loads on its own, skills fire when they're relevant, nothing else asks anything of you.

/done is what closes the loop. Skip it and the session's decisions live only in a transcript you'll never reopen — the next session starts from the same memory as the last one.

The failure mode is invisible, which is exactly why it has to become a habit: nothing breaks, nothing errors, memory just quietly stops growing. If you remember one thing from this README, make it this one.

Setup

/jarvis-setup

Opens a short-lived local server on 127.0.0.1 (random port, one-time token in the URL) and prints a link. In the browser you pick a hub directory (defaults to ~/jarvis-hub), toggle feature groups, and optionally point at an Obsidian vault. The hub is created and seeded with a starter MEMORY.md on save.

There is no daemon. The server exits when you close the tab (15s grace), hard-stops after 10 minutes idle, and dies with the terminal that launched it. It's a child process of your session, not something you have to remember to kill. Reopening is just /jarvis-setup again.

Until you run setup, the plugin stays quiet — one line at session start saying it isn't configured, nothing more.

Obsidian is recommended, not required. Everything works with a plain directory of markdown files. Obsidian adds browsing and wikilinks on top; the plugin never checks for it and doesn't degrade without it.

What this is NOT

• Not a clone of anyone's personal setup. Nothing here is tied to one person's projects, vault, or preferences. What ships is the generic engine plus system-level defaults — you fill in the memory yourself.
• Not a replacement for Claude Code's project memory. Different problem, and the two coexist fine.
• Not opinionated about how you work. Model-routing and session-hygiene defaults ship as documented suggestions with a stated rationale, never as the one correct way.

Reaching it from your phone

The default remote path is Claude Code's own Remote Control (claude remote-control + claude.ai/code on your phone). No bot token, no allowlist, nothing to configure.

Verified against the shipped binary (2.1.245), not the docs — so here's where it actually stops:

• Unavailable inside a cloud session and behind an enterprise gateway.
• Files the agent sends do not reach phone/web viewers.
• Only effortLevel and ultracode are changeable from the remote side.
• Needs the host awake with a live session. It's remote control of a running session, not a standalone remote agent.

A Telegram bridge is supported but opt-in and off the critical path — it needs a BotFather token and an allowlist, and two concurrent sessions polling the same bot will steal each other's messages. The setup page explains how to wire it when you want it.

Skills and licenses

Every shipped skill was checked against its primary source license before inclusion — not a blog post about the license, the actual LICENSE file.

• Vendored as-is (MIT): defuddle (Steph Ango / kepano, byte-identical to upstream) and qmd (Tobi Lütke). The bundled qmd is not current upstream — a trimmed fork of v2.0.0, while github.com/tobi/qmd is at v2.2.0 and has since added --full-path, line-slicing and qmd doctor. Treat it as a known-older excerpt.
• Ported and genericized from personal workflow skills, rewritten for a stranger's hub: done, memory-consolidation, wiki-article, learn, research-notebook, media-gen.
• Already generic, ships as-is: grill-me.

Recommended, but install them yourself

Left out for license reasons, not quality ones:

• llm-council — a question or decision run past five independent advisors, peer-reviewed, then synthesized. By Ole Lehmann, methodology credited to Andrej Karpathy. Its repo ships no LICENSE file, so there's no grant to redistribute it.
• nano-banana — image generation and editing, with a free tier. Licensed AGPL-3.0; bundling it would pull this entire plugin under AGPL.

Non-negotiables

• No secret is ever written into this repo, logged, or echoed back by the setup server. Keys go to your own config; the page shows a masked confirmation only.
• No telemetry. Nothing is sent anywhere.
• The local server is never a background process. Token-authenticated, loopback-only, and it kills itself.

Repo layout

.claude-plugin/       plugin + marketplace manifests
hooks/                SessionStart hook — injects hub MEMORY.md
commands/             /jarvis-setup
setup/                the ephemeral local server + its page
skills/               the 9 shipped skills
memory-template/      seed hub: MEMORY.md index + example memory files

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Security matters are in SECURITY.md.

License

MIT — see LICENSE.