The agent that builds itself.

Jarvis wakes on its own, learns from every issue, asks when it's stuck, and improves its own code — behind a seatbelt that never lets it ship a regression. Open source. Runs on your models (Claude, Codex, or fully local via Ollama).

curl -fsSL https://jarvisbot.app/install | bash

Beta. Default mode: shadow — it proposes and asks, never acts, until you let it. Run it on a box you trust.

SEE IT

▶ demo — streaming chat · the "Thought" block · /discover · web-verified answers (drop demo.gif / demo.mp4 here)
Chat — streaming replies, collapsible thinking, inline images, run a skill with /
Settings — connect a brain, route models with effort + context controls, add MCP servers

QUICKSTART

1. curl -fsSL https://jarvisbot.app/install | bashclones the core, sets up the venv + dashboard
2. Open the login link it printsit embeds your private access token — ./install.sh url re-prints it
3. Connect a brain in Settings → Brain & Modelsa Claude/ChatGPT login is cheapest; an API key or local Ollama works too

Runs on Linux. Optional Postgres + Redis (everything degrades gracefully without them).

WHAT IT IS

A sleeper that wakes

Each wake it rebuilds its world from durable memory, does one bounded thing, writes back what it learned, and sleeps. Runs forever on a finite model.

A swarm, not a script

It spawns cheap researchers on demand and coordinates them through shared memory — emergent, not hand-orchestrated.

Self-auditing

Before it trusts a conclusion it red-teams its own work. Before it changes its own code it must beat a tamper-proof fitness check it cannot fake.

It talks to you

When it genuinely can't figure something out, it asks — in the dashboard or on Telegram — like a colleague, and learns from your answer.

WHY IT'S DIFFERENT

Skeptical by construction

It won't accept a conclusion until that conclusion has survived an adversarial red-team — and it fails closed: an un-run check never counts as a pass. What it can't verify, it asks you.

Self-modification, caged

It can improve its own code — but only behind a tamper-proof fitness gate it can't fake, an independent red-team, and a one-command rollback. Off by default.

Your models, your machine

Claude/Codex via CLI, any OpenAI-compatible or Anthropic HTTP endpoint, or fully local via Ollama. Route a cheap model for grunt work and a frontier model for the main brain. Self-hosted, token-gated, secrets in a local vault. Nothing phones home.

Propose-only until trusted

Ships in shadow mode. Risky actions (deploy, infra, self-edit) are denied by default; you grant them deliberately and watch every tick in the Activity tab.

MODELS & TOOLS

Per-model controls, auto-detected

Jarvis reads what each model actually supports and shows only that: reasoning effort/thinking and the real context window — including Claude's 200K↔1M and per-model context sizing on Ollama. Switch it inline from the chatbox.

MCP, provider-agnostic

Add Model Context Protocol servers (local stdio or remote HTTP with one-click OAuth) and their tools are offered to whatever brain is answering — native tool-calls on HTTP backends, and the Claude CLI connects to the same servers itself.

Skills, one click

Shipped skills (web research, network discovery, deep code search) install with a click and run from chat with /. Paste your own, or just ask Jarvis to write one — it's auto-detected.

A dashboard, not a terminal

Streaming chat with a collapsible thought block, inline images, syntax-highlighted code, a knowledge base of what it learned, and live activity — all in a zero-dependency web UI.

THE LOOP

  wake → perceive (alerts · backlog · memory) → orient → decide (priority ladder)
       → think (the model) → ask if stuck → act (workers, propose-only) → reflect → sleep

  young Jarvis works on itself and learns first — your production work comes last, until it earns trust.

Memory: Redis (working) · Postgres (episodic) · optional semantic layer. Safety: every action class is opt-in; self-edits go snapshot → fitness → red-team → adopt or roll back.

OPEN CORE

The engine is public and environment-agnostic. Your infra lives in a private extras/ overlay — your config, your skills, your secrets — never in the core. Clone the core, add your overlay, run the installer. That's it.

FAQ

Is it safe to run?
It ships propose-only (shadow mode) — it won't change anything until you grant action classes. Run it on a box you trust; it's beta. Secrets stay in a local gitignored vault.
Does it phone home or need a cloud account?
No. It's self-hosted and runs on your own models — a Claude/ChatGPT subscription, an API key, or a fully local Ollama model.
What does it cost?
The software is free (Apache-2.0). You pay only for whatever model you point it at — and it routes cheap models for grunt work, a frontier model only for the main brain.
Can it really edit its own code?
Yes, but caged: a tamper-proof fitness gate it can't fake, an independent red-team, and a version archive to roll back to. It's off by default.
What do I need?
A Linux host and Python 3.10+. Postgres + Redis are optional (it degrades gracefully without them).
How do I help?
Star it, try it, and open an issue or PR. See the Contributing guide.