the agent that works while you don't
turn any model into an agent you can run unattended
terminal · cron · CI · telegram · discord
asks before touching anything real
curl -fsSL https://github.com/lvndry/jazz/releases/latest/download/install.sh | bash
a real session, not a screenshot
the five lanes count real concurrent work — mailbox, calendar and weather at once. and when jazz needs authority it stops and asks, wherever you are.
everywhere at once
The same agent, the same tools, the same memory — behind every front door. When a job needs your permission it asks you wherever you are, rather than stopping.
terminal · scripts · cron · CI · telegram · discord — the surface matrix
under the hood
Memory, skills and history are assembled per turn, then trimmed turn-aware — a message is never split in half. The window is a budget, and jazz counts it.
read the internals →18 providers behind one loop — OpenAI, Anthropic, Google, Mistral, Groq, OpenRouter… or ollama and llama.cpp with no API key at all.
read the internals →Every call is classified by risk before it executes. Reading your calendar just runs; writing to it stops and asks — and the ask reaches whatever surface you're on.
read the internals →Budget pressure, meltdown detection, and compaction: warn at 70%, compact at 80%, trim at 95. Watch the meter — this is the ladder, animated.
read the internals →the trust argument
A coding agent asks to edit a file you can revert. Jazz asks to send an email, move a meeting, post where other people read. There is no undo — so the approval card is the most consequential component in the product, and it travels to whatever surface you're on. Same facts, native dress: a band in the terminal, a message from a bot, a parked run in CI.
steal these
Production-ready recipes with install steps and risk tiers. Hover one to watch a condensed run.
Jazz is a free, open-source AI agent harness. It runs a general-purpose AI agent on your own machine that can use a terminal, read and write files, search the web, send messages, and work on a schedule — asking for your approval before anything consequential.
Yes. Jazz is MIT-licensed open source and costs nothing. You pay only your chosen LLM provider for tokens — or nothing at all if you run local models through Ollama or llama.cpp.
Yes. Jazz runs entirely on your machine and supports fully local models via Ollama and llama.cpp, so it works airgapped: no API keys, no telemetry leaving your disk, and an offline mode that disables every outbound request.
Eighteen providers behind one interface, including Anthropic, OpenAI, Google, Mistral, Groq, OpenRouter, and fully local options like Ollama and llama.cpp. Switch models without changing anything else about your agent.
A chatbot answers one prompt and stops. Jazz is an agent runtime: the same agent works in your terminal, in Telegram or Discord, in CI pipelines, and on cron schedules — unattended, with per-run approval policies controlling what it may do on its own.
Jazz treats safety as load-bearing: every tool declares an honest risk tier, mutating actions are gated behind explicit approval, external input is validated at the boundary, and scheduled runs get a per-workflow auto-approve policy you set.
read the docs — calm on purpose
curl -fsSL https://github.com/lvndry/jazz/releases/latest/download/install.sh | bash
or npm install -g jazz-ai · free to run:
OpenRouter's free router needs no card, and
ollama runs on your own hardware
MIT licensed · built in the open · /llms.txt