Claude Swarm
https://github.com/affaan-m/claude-swarm- Category
- AI Agents
- Rank
- No. 1096Tools index
Previous survey · No. 1102 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- affaan-m
- GitHub
- 353 stars
- Latest release
- v0.2.0
- Date
About
Multi-agent orchestrator for Claude Code that decomposes tasks across agents and visualizes everything in a rich terminal UI.
What it does
Give it a software change, and it first maps the work into ordered, partly independent jobs. Worker processes handle ready jobs concurrently, while later jobs wait for prerequisites. A final model pass assesses the combined results. During execution, the command-line dashboard reports task state, agent activity, spending, elapsed time, and conflicts. Runs are stored for later replay.
Why it's ranked here
The core workflow is unusually complete for an alpha package: planning, dependency scheduling, cost limits, retries, final review, and replay sit behind one command. The implementation also has meaningful weaknesses. Its advertised file protection appears ineffective because locks use task identifiers while conflict checks expect agent identifiers. A malformed quality-review response is treated as a pass. This makes Claude Swarm promising for supervised experiments, but risky as an unattended coding layer.
What's good
It covers practical operating concerns that many agent runners omit. Dependencies determine when work can start, concurrency is capped, each worker receives an allowed tool set, and failed work can retry. Dry runs expose the plan before execution. Demo mode exercises the interface without an API key. Cost reporting spans individual tasks and the whole run, while recorded event streams make agent activity inspectable afterward.
Tradeoffs
Real runs require an Anthropic API key and spend control is reactive: worker cost reaches the global total only after that worker finishes, so the stated ceiling can be exceeded. YAML support requires PyYAML, but the package does not declare it as a dependency. Although configuration describes custom agent models, tools, prompts, and connections, execution only applies its concurrency, budget, and planning-model settings. Conflict protection also appears broken by inconsistent lock ownership identifiers.
How to use it well
Use it for complex repository changes that split cleanly across files and have explicit dependency boundaries. Start with demo mode, inspect a dry-run plan, then execute under a conservative budget while watching the changes and final report. It best suits developers comfortable supervising autonomous edits and checking generated work themselves. Treat it as a terminal workflow, not an application integration layer: the supplied project exposes no HTTP API or MCP server.
Technical notes+
src/claude_swarm/decomposer.py asks the Claude Agent SDK for JSON tasks and falls back to one worker when parsing fails. src/claude_swarm/orchestrator.py schedules with AnyIO, fixes workers to Haiku, and records cost only after each result. Its _lock_files stores task.id before an agent exists, while _check_file_conflict resolves the stored value through self.agents, so planned-file collisions are unlikely to register. src/claude_swarm/quality_gate.py returns a passing QualityReport when review JSON cannot be parsed. src/claude_swarm/config.py imports YAML optionally, but pyproject.toml omits PyYAML and the CLI does not pass configured agent definitions into execution. src/claude_swarm/session.py writes metadata plus JSONL events under the user home directory.
Observed
- License
- MIT
- Primary language
- Python
- Runtime
- Python 3.11 or newer
- Packaging
- Hatchling-built PyPI package, installable with pip or as an editable source checkout
- Interface
- Click-based command-line interface with run, session listing, and replay commands
- Configuration
- Optional YAML topology format; PyYAML is not declared among package dependencies
- Tests
- The supplied repository includes pytest coverage for demo planning and retry configuration
Read from README.md, pyproject.toml, src/claude_swarm/ui.py, src/claude_swarm/cli.py, src/claude_swarm/demo.py, src/claude_swarm/types.py, src/claude_swarm/config.py, src/claude_swarm/session.py, src/claude_swarm/__init__.py, src/claude_swarm/decomposer.py, src/claude_swarm/orchestrator.py, src/claude_swarm/quality_gate.py, tests/test_demo.py, tests/test_retry.py.
What it can do
Decompose complex tasks into subtasks
Complex task description → Multiple smaller, manageable subtasks
Orchestrate multiple Claude agents
Task requirements and agent configurations → Coordinated multi-agent execution workflow
Distribute subtasks across available agents
Decomposed subtasks and agent pool → Task assignments to specific agents
Visualize agent workflow in terminal interface
Agent execution data and task states → Rich terminal UI displaying workflow progress
Monitor multi-agent task execution
Running agent processes and task status → Real-time execution status and progress updates
Aggregate results from multiple agents
Individual agent outputs and task results → Combined final result from all agents
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.