Skip to content

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

  • 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

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.
I want to… Go to
Get it running Installation
Follow a guided first run Tutorial
Point it at OpenAI / Anthropic / another provider Configuring a model
Understand trust levels and the destructive-action gate Concepts → Trust
Look up a command Command reference
See the separation guarantees proven Walkthrough
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.

Agent-to-agent collaboration and stronger execution isolation are still ahead; follow along in the repo's issues. For the precise scope of what "separate" guarantees today, see Concepts → What isolation means today.