<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> Claude Swarm orchestrates complex coding tasks by decomposing them into parallel subtasks, coordinating AI agents, and providing real-time visualization. It targets developers and power users seeking to automate multi-step code generation, refactoring, or analysis workflows, offering efficient, monitored, and cost-controlled execution.

How It Works

The system employs a multi-phase approach leveraging Claude models. Phase 1 uses Claude Opus 4.6 to analyze a task and codebase, generating a dependency graph of subtasks. Phase 2 executes these subtasks in parallel via the Swarm Orchestrator, utilizing Claude Haiku agents for cost-efficiency. Key features include dependency-aware scheduling, pessimistic file locking to prevent conflicts, and hard budget enforcement. Phase 2.5 involves an Opus 4.6 Quality Gate reviewing combined agent outputs for correctness. Phase 3 presents a summary of results, costs, and a session replay ID. This architecture mirrors a senior architect planning, junior engineers executing, and senior reviewing workflow.

Quick Start & Requirements

Install via pip: pip install claude-swarm. For development, clone the repo and run pip install -e .. Real usage requires setting the ANTHROPIC_API_KEY environment variable. A demo mode (claude-swarm --demo) visualizes the TUI without an API key. Requires Python 3.11+ and claude-agent-sdk (v0.1.35+).

Highlighted Details

• Dependency-aware scheduling ensures tasks execute only after their prerequisites are met.
• Pessimistic file locking prevents concurrent agent edits on the same file.
• Hard budget enforcement cancels remaining tasks when the specified USD limit is exceeded.
• Real-time cost tracking monitors per-agent and total expenses.
• Opus 4.6 Quality Gate reviews combined agent outputs for correctness and consistency.
• Strategic model selection uses Opus for planning/review and Haiku for cheaper worker agents.
• Session recording (JSONL) and replay functionality allow reviewing past executions.
• A Rich-based htop-style terminal UI visualizes agent progress, tool usage, and conflicts.

Maintenance & Community

Built for the Claude Code Hackathon (Feb 10-16, 2026). No explicit details on ongoing maintenance, core contributors, or community channels are provided in the README.

Licensing & Compatibility

• License: MIT.
• Compatibility: The permissive MIT license generally allows commercial use and integration with closed-source projects.

Limitations & Caveats

As a hackathon project, it may be experimental and lack extensive production hardening. Its effectiveness is contingent on Anthropic's Claude model performance and requires API access, incurring associated costs.