▞ jazz blog
Jazz vs the alternatives: which agent runtime fits where
If you’re evaluating AI agents, you’re really choosing between four shapes of product. They overlap less than the marketing suggests, and picking the wrong shape costs more than picking the wrong brand. Here’s the honest map.
The four shapes
Chat-first coding CLIs (Claude Code, Codex CLI, Gemini CLI, Aider) put a model in your terminal and let it edit code. They’re excellent at the job in their name: coding, interactively. What they generally don’t do is leave the terminal — running unattended on a schedule, answering you in Telegram while you’re out, or sitting inside a CI pipeline as a reviewer.
Hosted agent platforms (the cloud products from the big labs, plus no-code automation suites with LLM steps) run your prompts on their machines. That’s convenient until your agent needs your files, your git repos, your credentials, or your inbox — then you’re either shipping secrets to a third party or discovering the platform was never built for that.
Plain cron + scripts + an API key is the DIY shape. It works, right up until the run hits something unexpected — and then there’s no loop to recover, no context strategy, no approval gate, no record of what happened. You reinvent the harness badly, one bash script at a time.
Agent harnesses are the fourth shape: a general-purpose runtime whose whole job is making one model usable as an autonomous worker on your machine. Jazz is in this category by design.
Where Jazz differs
Jazz doesn’t try to be the best chat UI or the best autocomplete. It is built around three claims that the other shapes don’t make together:
One agent, every surface. The same configured agent works in your
terminal REPL, behind jazz run for scripts, in GitHub Actions, under
launchd/cron for scheduled runs, and in Telegram, Discord, or Slack. You
write the agent once; surfaces are deployment decisions, not rewrites.
Self-hosted by default. Jazz runs on your machine — laptop, home server, or CI runner. Pair it with Ollama or llama.cpp and it works fully airgapped: local inference, local telemetry, no outbound requests in offline mode. Nothing about your filesystem, credentials, or mail ever needs to leave the hardware you own.
Approval as a first-class mechanism. Every tool declares an honest risk tier; mutating actions are gated; scheduled workflows carry a per-workflow auto-approve policy that you set and can audit. The design goal is specific: an agent you can leave running unattended because you know exactly what it can’t do without asking.
The trade-offs, stated plainly
Choosing Jazz means choosing the trade-offs that come with self-hosting:
- You operate it. Updates, logs, and the machine it runs on are yours.
- There’s no hosted dashboard. Observability ships as local run records and OpenTelemetry export into whatever collector you already run.
- Chat surfaces require bringing your own bot tokens (Telegram, Discord).
- Interactive pair-programming inside an editor is not its center of gravity; tools built specifically for that will feel tighter in that one mode.
If those trade-offs bother you, use the tool shaped for you. If they sound like the point — an agent that lives on your machine, works while you don’t, and asks before touching anything real — that’s what Jazz is for.
Try it
curl -fsSL https://github.com/lvndry/jazz/releases/latest/download/install.sh | bashMIT licensed, single binary, no runtime dependencies. The quick start gets you to a working agent in a few minutes, and the playbooks has forkable recipes for inbox triage, PR watchdogs, and research digests.