// the find
unohee/OpenSwarm
OpenSwarm — Autonomous AI dev team orchestrator powered by Claude Code CLI. Discord control, Linear integration, cognitive memory.
OpenSwarm is an autonomous AI coding team orchestrator that wraps Claude Code CLI, Codex, GPT, OpenRouter, and local models (Ollama/LMStudio) into a Worker/Reviewer pipeline driven by Linear issues and controlled via Discord. It adds a vector memory layer (LanceDB), a code entity registry (SQLite), and a 'repo knowledge loop' where task outcomes are written back as embeddings to improve future prompts on the same codebase. Aimed at solo developers or small teams who want to delegate Linear ticket work to AI agents running unattended.
The hybrid benchmark result is concrete and credible — frontier model for read-only diagnosis, cheap model for implementation, official SWE-bench harness for grading, with raw result files checked into the repo rather than screenshots. The adapter abstraction is genuinely multi-provider: claude, codex, gpt, openrouter, and local models share a unified interface with per-role overrides, so you can mix haiku for workers and sonnet for reviewers without changing pipeline code. The memory retrieval formula (0.55 × similarity + 0.20 × importance + 0.15 × recency + 0.10 × frequency) is explicit and tunable rather than a black box. The TUI chat works with zero config — `npx @intrect/openswarm` with no arguments drops you into a Claude session immediately, which is a sensible default for casual use.
The SWE-bench claim is 3/3 attempted instances, which is far too small a sample to benchmark anything — it's three cherry-picked pylint issues, not a statistically meaningful result, and the repo is honest enough to include all the failed intermediate attempts if you look at the benchmarks/results directory. The full daemon setup has a heavy dependency surface: Claude Code CLI authenticated, Discord bot with specific intents, Linear API key, GitHub CLI, plus optional Codex/OpenRouter tokens — getting all of these configured correctly is a real integration tax that the README underplays. The memory system writes task outcomes to LanceDB per-repo but there's no described mechanism for pruning stale or wrong knowledge entries, so a misguided agent run could poison future prompts on that repo indefinitely. GPL-3.0 license will stop most commercial teams from adopting it without legal review.