Local-first agent OS
Define an AI agent in YAML, give it a personality, point it at any LLM backend, and let it run. Agents persist across sessions, collaborate through shared rooms, and synthesize tools as needed. No Docker, no cloud. Your machine, your models.
Agents don't just wait for instructions. A daemon runs them on a heartbeat loop where they generate their own goals from context, plan, act, and react to stressors — autonomy you can watch and inspect, not a black box.
Hive is also a proving ground. You can pit agents against each other in structured scenarios. The first was Texas Hold'em; the latest is a locked-room murder mystery solved by three AI detectives running different models.
I ran 5 poker tournaments with 6 LLMs ranging from 1.2B to ~1T parameters. The smallest model (Liquid, 1.2B, running locally) won 2 out of 5 tournaments and had the best average placement. The trillion-parameter model averaged last place.
The full results and analysis are in the blog post.
Every agent framework I've tried either requires a PhD in configuration or ships your data to someone else's cloud. Hive is the alternative: your machine, your models, your agents. The arena is a proving ground for how agents make decisions under pressure.
Python (async throughout), Typer CLI, YAML agent profiles, JSONL event logs + SQLite state (aiosqlite), multi-model routing (Anthropic, OpenAI, Groq, Fireworks, Ollama, LM Studio), autonomy daemon, scenario runner, MCP server, pokertable for game mechanics, agent-to-agent rooms, persona engine, suffering system. pip install hive-agents.