- CollectionAgent Harnesses & Operator Surfaces
- Classificationagent-native
- LicenseApache-2.0
- MCPUnavailable
- Agent SkillUnavailable
- InterfacesCLI · http
- Verified2026-10-27
What it does
Reef connects agent inference, feedback, learning, and versioned delivery. Recipes can train model weights (e.g. with Slime/SGLang) or improve an agent’s harness (prompts, rules, skills) without local training GPUs. Learning cycles: Serve → Observe → Grow → Commit, with version history for accepted updates.
Agent-facing interfaces
- Interaction pattern: Local/self-hosted Python infra (CLI)
- Catalog protocols: CLI, http
- MCP: ⚠️ No MCP server in the verified materials — primary surface is the reef CLI / Python package (serve, train, evaluate, artifact delivery).
- Agent Skill: ⚠️ No official AgentSkills bundle located; integration is the reef-infra Python package and CLI.
- Onboarding type:
cli
How agents onboard
Interaction pattern: Local/self-hosted Python infra (CLI)
uv venv && source .venv/bin/activate
uv pip install reef-infra
python3 -c "import reef; print(reef.__version__)"
# minimal inference serve example from upstream README:
uv run reef serve --inference.model-path Qwen/Qwen2.5-1.5B-Instruct
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| PyPI | uv pip install reef-infra |
| CLI | reef serve, training/eval modules under the package |
| Docs | https://reefinfra.ai/docs/getting-started/quickstart/ |
| Source | https://github.com/Human-Agent-Society/reef |
Verification
Last verified 2026-10-27.
Sources:
- https://github.com/Human-Agent-Society/reef