GET STARTED

Up and Running in Minutes

A curl command, a scope choice, and you're in.

Install

curl -fsSL https://raw.githubusercontent.com/timgranlundmarsden/claude-agent-flow/main/install.sh | bash
Run from inside your project's git repository. The installer prompts you to choose a scope.

Scope-specific commands:

# Plugin only (lightest)
curl -fsSL https://raw.githubusercontent.com/timgranlundmarsden/claude-agent-flow/main/install.sh | bash -s -- --scope plugin
# Plugin + GitHub Actions
curl -fsSL https://raw.githubusercontent.com/timgranlundmarsden/claude-agent-flow/main/install.sh | bash -s -- --scope plugin+github
# Sandbox (fully self-contained)
curl -fsSL https://raw.githubusercontent.com/timgranlundmarsden/claude-agent-flow/main/install.sh | bash -s -- --scope sandbox

Scopes

plugin

Lightest
Installs as a Claude Code Plugin — the 12-agent team, commands, and skills are managed by Claude Code's plugin system and kept up to date automatically. Your repo gets only the connecting config: CLAUDE.md patch, settings.json, .gitattributes, .mcp.json, Backlog.md init.
Best for: Teams managing their own CI/CD

sandbox

Self-contained
Everything fully vendored into your repo — agents, commands, and skills live alongside your code with no external plugin dependency at runtime. Full pipeline works inside sandboxed environments including Claude Code web (claude.ai/code)
Best for: Claude Code web or air-gapped environments
Two ways to get sandbox mode:
A
Existing repo — use the curl command above with --scope sandbox
B
New project — clone the repo, open Claude Code in the folder, run /install
git clone https://github.com/timgranlundmarsden/claude-agent-flow
Sandbox scope is auto-selected — no prompt.

Prerequisites

  • Claude Code CLI installed
  • Git repository initialised in your project
  • Node.js 18+ (for MCP servers and Backlog.md)

Your First Task

1
Open Claude CodeOpen Claude Code in your project root.
2
Plan your feature
/plan I want to build a CV web page
3
Build your featureOnce planning is complete:
/build @my-plan.md
4
Review the PRThe pipeline creates a PR — reviewed, tested, and documented — ready for your sign-off.
💡
Tip: unattended mode — For fully autonomous multi-agent runs with no per-tool prompts, add "defaultMode": "bypassPermissions" to your .claude/settings.local.json (gitignored — applies to your machine only). Only recommended for trusted repos where you're comfortable with autonomous tool use.
📋
Your stack, detected automatically — On first pipeline run, the explorer agent scans your project and creates TECHSTACK.md with your technology profile. Agents use this to know your test runner, linter, and conventions — no hardcoded assumptions. Edit it at any time; your changes are always authoritative.

What You Get

12 Agents

orchestratorCoordinates the team
explorerAnalyzes codebase
architectDesigns system structure
ideatorGenerates creative solutions
researcherGathers information
frontendUI/UX implementation
backendServer-side logic
storageDatabase & persistence
testerQuality assurance
criticAdversarial review
reviewerFinal code review
authorDocumentation

13 Commands

/buildFull pipeline
/planPlanning session
/reviewCode review
/exploreCodebase analysis
/rebaseGit workflow
/installSystem setup
/diagnosticHealth check
/helpDocumentation
/backlog-listTask management
/backlog-statusTask management
/check-prPR validation
/external-reviewAI code review
/token-analyserCost tracking

18 Skills

agent-development
ascii-box-tables
backlog-md
backlog-tpm
brainstorming
command-development
external-code-review
frontend-design
hook-development
mcp-integration
playwright-cli
playwright-cli-helpers
plugin-settings
plugin-structure
skill-development
token-analyser
ways-of-working
agent-flow-init-check
log-visualiser-converter