Skip to content

Prometheus v0.9.0

Pre-release
Pre-release

Choose a tag to compare

View all tags
@github-actions github-actions released this 24 Sep 02:15
· 302 commits to main since this release
v0.9.0
bfa2701
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The first release of Prometheus on PyPI.

Prometheus is a self-hosted agent harness for local LLMs. It's an always-on daemon that makes open models reliable at calling tools, and it keeps your memory and telemetry on your own hardware. It pairs with Beacon, a free desktop app for macOS and Linux, and with an iOS client.

What it does

  • Reliable tool calls on open models. A Model Adapter Layer checks every tool call, repairs common mistakes (misspelled tool names, JSON wrapped in markdown, wrong types), and when a call still fails, tells the model exactly what was wrong. On llama.cpp it can constrain output with a grammar.
  • Local first, cloud welcome. llama.cpp, Ollama, LM Studio and vLLM on your own machines; OpenAI, Anthropic, Gemini, xAI, DeepSeek, Kimi, GLM and MiMo in the cloud, switchable per chat. Name several local GPU machines and point any chat at one of them.
  • Always on. Telegram, Slack and Discord gateways, scheduled jobs, and background tasks that survive restarts.
  • Memory you can read. Plain MEMORY.md and USER.md files, facts mined from your conversations, and context compression that summarises long sessions while keeping the originals searchable and recoverable.
  • Coding runs that end in a branch. Point it at a repo and a test command. It works in a sandboxed clone until the command passes, then leaves a branch for you to review. It never merges or pushes.
  • A security gate. Trust levels with approval prompts, workspace boundaries, audit logs, and API keys kept out of the environment that tools run in.
  • Telemetry that stays home. Every tool call and token count is logged to SQLite on your own disk and sent nowhere.
  • Open interfaces. Token-authenticated REST and WebSocket APIs, plus an OpenAI-compatible /v1 endpoint.

Install

pip install 'oara-prometheus[full]'
oara setup
oara daemon

Python 3.11 or later. If something's off, oara doctor checks every subsystem and suggests a fix for each failure.

Prometheus is in active development, so expect rough edges. On Intel Macs with Python 3.14, 0.9.0 doesn't install because a voice dependency has no wheel there. 0.9.1 takes that dependency out of the base install by making voice an optional extra.

Assets 4
Loading