#2727th of 42 in Agent platforms
reef
Serves agent traffic, collects feedback, and publishes versioned weight or harness updates
Documentation ↗
(opens in a new tab)Website ↗
(opens in a new tab)Repository on GitHub ↗
(opens in a new tab)
- Stars
- 7.7k
- License
- Apache-2.0
- Last commit
- Oct 2026
- Last release
- Oct 2026
- Language
- Python
182 on 2 Sept 20267,724 on 21 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Reef is a Python server that exposes OpenAI- and Anthropic-compatible endpoints, records each interaction, and matches later score or feedback reports to it by receipt header. A configured recipe then either trains model weights (Slime with SGLang) or rewrites the agent harness (prompts, rules, skills), and accepted updates are committed to a version history without restarting the server.
Who it is for: Teams building agents that learn from feedback via RL or harness evolution
Strengths
- Harness-evolving recipes (Reefine) need only a model endpoint, no training GPUs
- Serves /v1/chat/completions, /v1/responses and /v1/messages while updates roll in live
- Versioned artifact history with candidate evaluation; rejected candidates leave the current release serving
- Recipe catalog with measured benchmarks (AIME 2025, IMOAnswerBench, Terminal-Bench, WildClawBench)
Weaknesses
- Weight training needs a supported GPU stack; exact GPU requirements are only in linked docs
- Requires Python 3.12+ and git-lfs; no Docker or compose files provided
- Most recipes live in the repo's recipes/ cookbook and are not shipped in the wheel
- Reefine proposer sandbox needs Linux with bwrap and pasta; pi runs commands unsandboxed
What it needs
- GPU optional
- Docker
- Needs git-lfs, Slime, SGLang, uv
- Models: Qwen/Qwen2.5-1.5B-Instruct, Ollama-served models (gemma4:26b), OpenAI-compatible upstream providers
- port 8900
Also in Agent platforms
See all 42| Rank | Project | Score |
|---|---|---|
| 1 | LangflowVisual builder for AI agents and workflows, exposed as API or MCP server | 81 out of 100 |
| 2 | daily_stock_analysisDaily LLM stock analysis for A-share, HK, US and other markets, pushed to chat apps | 81 out of 100 |
| 3 | resume-matcherTailors resumes and cover letters to job descriptions using local or hosted LLMs | 81 out of 100 |
| #4 | vibe-tradingFinance research agent for backtests, market data and broker connectors | 78 out of 100 |
| #5 | deer-flowAgent harness that runs sub-agents, memory and sandboxed skills for long tasks | 77 out of 100 |