Flow-Claude is for Claude Code users working on long development tasks.
- Break down requirements into parallel tasks,
- execute them simultaneously without stopping,
- auto-commit every change to Git,
- and merge results automatically.
Niu, B., Song, Y., Lian, K., Shen, Y., Yao, Y., Zhang, K., & Liu, T. Flow: Modularized Agentic Workflow Automation. ICLR 2025
We design Flow-Claude as a lightweight tool that lives within the Claude CLI. As the Claude code model evolves, the benefits of the framework will also continue to grow.
The framework itself should not become a blocker for future updates of the Claude model. Therefore our primary focus is communication efficiency and parallelism with minimum constraints.
Every new design should smoothly support the Claude CLI.
Submit a GitHub issue or contact yu.yao@sydney.edu.au
Claude Code must be installed first:
npm install -g @anthropic-ai/claude-code
See Claude Code Setup for more details.
Install Flow-Claude From PyPI
pip install flow-claudeVerify Installation:
flow --helpOn Windows, if flow is not in your PATH, use:
python -m flow_claude.commands.flow_cli --helpNavigate to your git repository and initialize Flow-Claude:
cd /path/to/your/project
flowOn Windows, if flow is not in your PATH, use:
cd /path/to/your/project
python -m flow_claude.commands.flow_cliWhat happens during initialization:
- Creates
flowbranch from your main branch - Creates
.claude/directory with skills, commands, agents - Creates/updates
CLAUDE.mdwith Flow-Claude instructions - Commits all changes to
flowbranch
claudeIf you want to avoid frequent permission approval, use:
claude --dangerously-skip-permissionsBecause the current Claude model does not reliably invoke Skills automatically, you need to append a reminder at the end of your prompt.
For example:
Build a REST API for blog posts with CRUD operations. Remember to use your SKILLs.
When you return to a project after a break or session interruption, prompt Claude to understand the current state before continuing.
Step 1: Use the resume command to restore your previous conversation:
/resume
Select the past conversation from the list.
Step 2: Prompt Claude to check the current state and continue. For example:
Check the current plan and task branch status, then continue the unfinished task. Remember to use Skills.
If you make local changes on the flow branch and want workers to use them, commit your changes first:
git add . && git commit -m "Your commit message"Workers pull from the flow branch, so uncommitted changes won't be visible to them.
After a plan is finished and all tasks passed, you may want to clean up the created branches. Use a prompt like:
Delete the plan/<plan-name> branch and all its related task branches.
For example: Delete the plan/blog-api-crud branch and all its related task branches.
| Command | Description |
|---|---|
\auto |
Toggle autonomous mode (ON = no approval needed) |
\parallel N |
Set max parallel workers (1-10, default: 3) |
The main agent can automatically determine which MCP tools and agent skills the worker agents require.
Flow-Claude uses MCP in the same way as Claude Code, except that MCP servers must be installed inside your project directory.
For detailed MCP setup instructions, see:
Example: Adding the Playwright MCP server
claude mcp add playwright -- npx --scope project @playwright/mcp@latest # Run this command inside your project directoryOr manually add to .mcp.json in your project root folder:
{
"mcpServers": {
"playwright": {
"type": "stdio",
"command": "npx",
"args": ["@playwright/mcp@latest"]
}
}
}Flow-Claude uses Skill in the same way as Claude Code, except that Skill must be added inside your project directory.
For detailed Skill adding instructions, see:
Example: Creating Skills
- Created as
SKILL.mdfiles in specific directories (.claude/skills/)
.claude/skills/processing-pdfs/
└── SKILL.md
MIT License - see LICENSE for details.
Ready to supercharge your development?
pip install flow-claude && cd your-project && flow

