Asterism documentation¶
Run many distinct AI agents from one local install — each with its own soul, memory, secrets, skills, workspace, event log, and autonomy level.
Nothing leaks between them.
New here? Start with the getting-started tutorial — a ~15-minute walk from install to a working agent that writes a file, pauses before deleting one, and remembers what you approve.
Getting started¶
- Installation — install Asterism on Node, Bun, or Deno, initialize a workspace, and configure a model.
- Tutorial — the gentle, end-to-end on-ramp: create an agent, run it, watch the destructive-action gate fire, and approve a memory.
- Concepts — agents, souls, roles, trust, memory and recall, skills, secrets, standing objectives, and working notes — and exactly what "separate" means today.
Guides¶
- Working together — open a channel between two agents, choose what it is for, and withdraw it.
- Dashboard — watch and steer every agent in one live terminal view.
- Chat channels — reach one agent from a Telegram or Discord chat.
- Run as a service — keep an agent running in the background, started by your OS.
- Run in a container — package the same runtime to run on any container host.
- Local HTTP endpoint — serve one agent over HTTP, with the same guarantees as the command line.
Reference¶
- Command reference — every command, option, and its output.
Deep dive¶
- Five-claims walkthrough — the canonical demo, with the separation and trust guarantees proven end to end. The skeptic's version of the tutorial.
- Threat model — what the runtime enforces and by what mechanism, each claim carrying the test that proves it — and, just as plainly, what today's boundary does not contain.
Quick links¶
| I want to… | Go to |
|---|---|
| Get it running | Installation |
| Follow a guided first run | Tutorial |
| Point it at a model — hosted, or one on your own machine with no key | Models and providers |
| Understand trust levels and the destructive-action gate | Concepts → Trust |
| Have two agents work together | Working together |
| Let an agent call an API with a stored credential | Command reference → api |
| Choose which tools an agent has | Command reference → capabilities |
| Look up a command | Command reference |
| See the separation guarantees proven | Walkthrough |
| Know exactly what is enforced — and what is not | Threat model |
| Call an agent over HTTP | HTTP endpoint |
| Watch and steer every agent at once | Dashboard |
| Keep an agent running in the background | Run as a service |
What Asterism does today¶
Asterism runs distinct agents from one local install — each with its own memory, secrets, skills, workspace, and autonomy, with a destructive-action gate that holds at every trust level. On top of that core it adds governed learning — how an agent improves with you in control: structured recall, earned per-capability trust, reviewed reflection you can put on a schedule, standing objectives, and the agent's own working notes. You can drive an agent from the live dashboard, a Telegram or Discord chat, an HTTP endpoint, or a background service, and run the whole thing in a container.
Agents can also work together without giving up any of that: you open a channel between two of them by hand, choose what it is for, and only that ever crosses — see Working together.
Stronger execution isolation is still ahead; follow along in the repo's issues. For the precise scope of what "separate" guarantees today, see Concepts → What isolation means today.