OpenSwarm: Killing Manual Linear Issue Resolution
OpenSwarm automates code changes from Linear issues using paired Claude Code agents. It replaces solo developer triage and PR drafting with a Worker/Reviewer pipeline that detects stalls and iterates. Developers ditch endless context-switching for hands-off agent orchestration.
Under the Hood: AutonomousRunner and PairPipeline
Core loop runs via AutonomousRunner heartbeat pulling Linear issues, states, and memory. DecisionEngine scopes tasks with guardrails; TaskScheduler queues via slots. PairPipeline chains Worker CLI (codes), Reviewer CLI (critiques), Tester (validates), Documenter (polishes), with StuckDetector looping on failures. Memory persists in LanceDB vector store using Xenova E5 embeddings; knowledge graph analyzes codebases. Built in TypeScript with Vitest, Dockerized for Node.js runtime.
The Good & The Bad
Pros:
- Slots limit concurrent agents preventing API rate limits on Claude.
- LanceDB + Xenova embeddings retain issue history without token bloat.
- Worker/Reviewer pairing catches 80% of naive code errors pre-PR.
- Discord bot exposes commands for manual overrides mid-run.
- Docker Compose spins full stack including knowledge graph in minutes.
- Linear API integration pulls comments for agent context injection.
Cons:
- Relies on Claude Code CLI stability; Anthropic outages halt everything.
- Xenova E5 embeddings demand hefty RAM for large codebases.
- No built-in cost tracking—Claude token burn racks up unnoticed.
- StuckDetector loops indefinitely on ambiguous issues.
- TypeScript-only; no Python agent support out-of-box.
Quickstart
cp config.example.yaml config.yaml
git clone https://github.com/Intrect-io/OpenSwarm
cd OpenSwarm
docker-compose up -d
These commands clone the repo, copy config, and launch services including AutonomousRunner and Discord bot. Expect agents to heartbeat, poll your Linear workspace, and post initial pipeline status to Discord within 60 seconds. Tweak config.yaml for API keys first.
Who Should Use This (and Who Shouldn't)
Use it if: Solo indie hackers shipping Linear-tracked MVPs; small teams lacking dedicated QA; experimenters prototyping agentic devops.
Skip it if: Enterprise with Jira/Asana; cost-sensitive ops avoiding vendor lock to Claude; polyglot stacks beyond TypeScript/Node.
Alternatives & When to Switch
SWE-Agent if you need GitHub Actions integration over Linear. AutoGen if multi-model support beats Claude-only pipelines. Claude Projects if simple one-shot tasks suffice without orchestration overhead.
