Jazz Documentation

Jazz runs your AI agent everywhere you are — terminal, script, cron, CI, chat. Any model, including local ones. These docs are organized by what you’re trying to do.

Start here

I want to…	Go to
Install it and see it work	Quick Start
Know where it can run	Where it runs
Copy a finished thing	Playbooks
Understand a concept	Concepts
Look up a flag or tool	Reference
See how it works inside	Internals
Understand the interface design	Design

Sections

Start — get running

Quick Start — install, configure a provider, first answer
Creating Agents — build an agent for a job
Creating a Telegram or Discord bot — bot token to a working agent in your chats
Airgapped & Self-Hosted — fully offline with Ollama or llama.cpp
Observability — telemetry to your own OpenTelemetry collector or Langfuse

Use cases — concrete jobs Jazz is good at

One agent, many front doors. Start with the matrix below.

Headless — the jazz run contract: stdout/stderr, JSON envelope, per-chat memory, live events
Chat platforms — Telegram and Discord (shipped), Slack / Google Chat (bring your own bridge)
CI/CD — PR review with inline comments, the /jazz assistant, release notes
Scheduled — launchd / cron, catch-up, unattended safety

Concepts — the building blocks

Agents · Personas · Skills · Tools · Workflows · Scheduling · Agent-to-agent · Webhooks

Walkthroughs

Concrete sessions from ask to artifact: setting up peers, deep research into Obsidian, git history surgery, security scans, and more.

Playbooks — copy-pasteable recipes

Production-ready workflows with install steps and risk tiers: inbox triage, PR watchdog, competitor watch, tech-debt radar, research digest, CI reviewer, release notes.

Integrations — connect things

LLM Providers — 18 of them, including local
MCP Servers · Web Search · Email & Calendar

Reference — look it up

CLI · Configuration · Tools · Workflow frontmatter

Internals — how it works

Agent loop — iterations, budget pressure, meltdown detection
Context management — token counting, trimming, compaction
Tools & approval — risk tiers, two-phase execution
Sub-agents · Skills loading · Providers & models
Evals — measuring whether a harness change actually helped
Design decisions — every harness choice and what it trades away
Code map — for contributors

Design — how the interface is built

The terminal interface: the mark, the single-column layout, the six-hue palette, the activity indicators, and the approval card. Includes the two rules that decide every case not covered explicitly, and how the same design holds over SSH, in a cutting-edge terminal, and with no terminal at all.

Security

The threat model, the approval tiers, hardening for unattended and chat-facing deployments, and how to report a vulnerability. Lives at the repository root.

Help

Discord — fastest way to get an answer
GitHub Discussions — ideas and questions
Issues — bugs and feature requests
CONTRIBUTING.md — contributor guide