Horae¶
Horae is a predictable, self-hosted personal agent. You declare a pipeline in YAML
(collect → transform → llm → act → notify); Horae runs it on a schedule, and every change it
makes can be previewed, journaled and undone.
Early software
Horae is pre-release. This site describes the code on main; the design decisions it
points to are provisional until the maintainer accepts them, and some features are
still planned (the pages say which). Check the roadmap.
Where to start¶
| I want to... | Read |
|---|---|
| understand the vocabulary | Concepts |
| run Horae with Docker or on Coolify | Quick start |
| write a plugin | Writing a plugin |
| know what Horae defends against | Security model |
| look up a command | CLI reference |
| operate a running instance | Docker Compose, Coolify, backups |
The ideas in one minute¶
- Pipelines, not agent loops. The control flow is written down, never chosen by a model (ADR 0001).
- One bounded LLM step. A skill, an output schema, a model alias and a budget, and no tools.
- Plugins propose; the core applies. A plugin returns intents; the core validates them, shows a dry-run diff, journals them and applies them (ADR 0002).
- Declared permissions. Hosts, secrets, paths and intents are approved at install and pinned
in
horae.lock. - Nothing updates itself. Plugins come from git at a pinned commit (ADR 0004).
- Bring your own model: the Anthropic API, your own
claude -p, OpenAI-compatible endpoints or Ollama.
Source and issues: github.com/horae-dev/horae.