hive

Hive is an open-source runtime for running many agents on business work. You hand a task to a long-lived queen agent with a role and memory; she does the first unit herself to learn the method, saves it as a reusable skill, then fans the rest out to parallel worker clones. Every worker writes one row into a shared SQLite tracker, so the queen checks progress with SQL and can resume where it stopped.

There is one execution primitive, the agent loop, and orchestration is a tool call (run_worker, run_playbook) rather than a compiled graph. Four workers run at a time by default, each under a hard tool-call budget. The tracker on disk, not a context window, keeps count of what is done, and re-running a playbook picks up the rows that still need work, with retries, backoff and a dead-letter list.

It ships thirteen persona queens, markdown memory benchmarked on LongMemEval, in-process tools (shell, files, search, scraping, charts), a browser bridge for your own Chrome, Agent Skills support and any MCP server via hive mcp add. State is saved to disk so agents resume after a crash, and it runs on most providers through LiteLLM, including local Ollama.