EverMind open-sources Raven: a harness-of-harnesses DAG over Claude Code, Codex, and Pi; ★5097 Apache-2.0

Why it matters

Terminal coding agents (Claude Code, Codex, Hermes, Pi, …) leave decomposition, dependency edges, artifact handoffs, and recovery across panes and human memory. EverMind’s open-source Raven positions itself as a harness of harnesses: a Host Agent that compiles goals into a validated DAG over built-in specialists and third-party harnesses—not another single-model chat shell.

Verified now: GitHub EverMind-AI/Raven ★5097 / 145 forks; Apache-2.0; latest release v0.2.3 (2026-09-27); site raven.evermind.ai; arXiv 2609.33439. Labeled pre-alpha. Active Oct-3 commits (including self-config) show the orchestration layer is still moving.

Architecture

The host builds a typed dependency graph; nodes bind registry backends (in-process Raven loop, CLI, ACP, or OpenAI-compatible API). Admission checks run before dispatch; a ready-set scheduler, completion judge, and exception path let the host continue, abandon, or replan. Built-ins: Raven-Research / Code / Design / Oncall. Third parties connect via ACP presets (Claude Code, Codex, OpenCode, Hermes, OpenClaw, Pi, Grok Build, Copilot, … per README). “RSI” here means proposing/validating changes to Raven’s own harness modules—not rewriting vendor CLIs. Experimental Curator ships in-repo under experimental/, not in the installed wheel (per v0.2.3 notes).

Evidence

Official eval is layered. Paper abstract: on MAOB, Raven ranks first on all four graph metrics under both tested backbones, with Exact Match +10.4 / +10.5 pp vs the strongest baseline (planning quality before worker execution). Product page: Raven-Research 76.5% on DeepResearch Mixed. Coding/design/oncall figures exist in README—no invented SWE numbers here; re-check figure captions and the tech report. Treat saturated-bench deltas skeptically and validate on your own tasks.

Try it

curl -fsSL https://raven.evermind.ai/install.sh | bash (Linux/macOS/WSL2); Windows via install.ps1. Needs uv + Node 22 via installer. Then raven web or raven tui. Docs: Quick Start. Start with a small Research→Code graph in an isolated repo before attaching third-party ACP agents; keep Curator off critical trees until you accept pre-alpha risk.