Inspectable building blocks for teams that want more control than a closed agent platform allows.
Reef captures agent usage, feedback, and evaluation, then versions improvements to model weights or agent harnesses through governed recipes and reviewable releases.
Why it made this list: Governed improvement loops for agents in production.
Self-hosted web/mobile personal agent with persistent Chromium, optional Linux workspace, durable tasks, approvals, goals, takeover console and receipts.
Why it made this list: Persistent browser and terminal work with takeover controls.
Runtime for agent fleets in kernel-isolated sandboxes with declarative policy, credential injection, observability and runtime checks.
Why it made this list: A security-minded runtime boundary for autonomous agents.
Local 0.6B decision model returns calibrated distributions for boolean, choice and score questions without text decoding.
Why it made this list: Local agent memory with an explicit evidence trail.
Self-hosted harness joins Claude Code, Codex and Pi into persistent named teams with durable roles, queues, direct messages, TUI, MCP and tmux workspaces.
Why it made this list: A portable runtime for agent tools and execution.
Agent memory layer with retain, recall and reflect, mental models and knowledge pages; self-hosted, embedded or managed.
Why it made this list: Long-term agent memory with strong retrieval primitives.
