Autonomous multi-agent code generation and PR automation for 13 AI coding assistants. Scaffold a structured workspace where specialized agents (orchestrator, architect, developer, reviewer, DevOps, designer) operate in a gated pipeline (PLAN → BUILD → TEST → VERIFY → SHIP) to drive software from backlog to merged PR autonomously — without context collapse, without code review bottlenecks, without hand-holding. You set direction; the agents handle execution.
Tags: AI agent orchestration · multi-agent AI · autonomous code generation · Claude Code · AI workflow automation · code review pipeline · agent routing · multi-LLM support · Claude Sonnet · OpenAI Codex · Gemini CLI
Which command do I need?
| Situation | Command |
|---|---|
| Starting a new project | multiagent-setup new <name> |
| Adding to an existing git repo | multiagent-setup init . |
macOS — Homebrew (no .NET required):
brew install Neftedollar/multiagent-template/multiagent-setup
multiagent-setup new MyProject # new workspace
multiagent-setup init .# add to existing repoOne-liner bootstrap (installs all deps including Homebrew/winget, agent CLI, and creates the workspace):
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/Neftedollar/multiagent-template/main/bootstrap.sh | bash -s -- MyProject
curl -fsSL https://raw.githubusercontent.com/Neftedollar/multiagent-template/main/bootstrap.sh | bash -s -- MyProject --provider gemini
# Windows (PowerShell)
irm https://raw.githubusercontent.com/Neftedollar/multiagent-template/main/bootstrap.ps1 -OutFile bootstrap.ps1
.\bootstrap.ps1 MyProject
.\bootstrap.ps1 MyProject --provider codexNuGet (.NET 10+ required):
dotnet tool install -g multiagent-setup
multiagent-setup new MyProject # Claude (default)
multiagent-setup new MyProject --provider nessy # Nessy CLI (Claude alias)
multiagent-setup new MyProject --provider codex # OpenAI Codex
multiagent-setup new MyProject --provider qwen # Qwen Code
multiagent-setup new MyProject --provider cursor # Cursor IDE
multiagent-setup new MyProject --provider windsurf # Windsurf IDE
multiagent-setup new MyProject --provider copilot # GitHub Copilot
multiagent-setup new MyProject --provider gemini # Google Gemini CLI
multiagent-setup new MyProject --provider cline # Cline (VS Code)
multiagent-setup new MyProject --provider aider # Aider AI pair programmer
multiagent-setup new MyProject --provider continue# Continue.dev (VS Code / JetBrains)
multiagent-setup new MyProject --provider roo # Roo Code (VS Code)
multiagent-setup new MyProject --provider kiro # Amazon Kiro (VS Code)
multiagent-setup new MyProject --provider all # all providers at once# Choose a workspace template (saas, oss, internal, or default)
multiagent-setup new MyProject --template saas # SaaS product (user-impact gates, feature flags, SLO review)
multiagent-setup new MyProject --template oss # Open source (CHANGELOG, semver, backward-compat gates)
multiagent-setup new MyProject --template internal # Internal tools (lighter pipeline, no external gates)# Add workspace files to an existing git repo (does not touch your code)
multiagent-setup init .# current directory
multiagent-setup init ./my-repo # specific directory
multiagent-setup init . --template saas # with template# Add a provider to an existing workspace (no need to recreate)
multiagent-setup add-provider cursor
multiagent-setup add-provider gemini --force # overwrite existing files
multiagent-setup add-provider all # add all providers at once# Manage providers
multiagent-setup list-providers # show installed vs. available providers
multiagent-setup remove-provider cursor # cleanly remove a provider and its files# Update an existing workspace to the latest templates
multiagent-setup update # skip already-customised files
multiagent-setup update --force # overwrite everything (CLAUDE.md preserved)Then start working:
cd MyProject
claude # or: nessy / codex / qwen-code / gemini (terminal agents)# or open in Cursor / Windsurf / VS Code (IDE agents)
/orchestrator Implement user authentication with JWTWhat happens next:
- Orchestrator reads
docs/process.mdand selects roles dynamically - Creates a plan and asks:
APPROVED / NEEDS WORK? - Proceeds through PLAN → BUILD → TEST → VERIFY → SHIP
- Opens a PR on GitHub when done — no manual intervention needed
Interactive mode: Claude asks for gate approval at each step.
Autonomous mode (claude -p): runs fully headless, escalates blockers to GitHub Issues.
What you get in 5 minutes:
- A workspace where agents run
PLAN → BUILD → TEST → VERIFY → SHIPautonomously - 20+ specialist roles installed in
~/.claude/commands/— orchestrator, architect, developer, reviewer, DevOps, designer, and more - Safety hooks active: dangerous commands blocked, commits enforced as conventional, files auto-linted on save
- One human touchpoint: review and merge the PR when the pipeline finishes
The problem: Asking a single AI agent to be architect, developer, reviewer, and DevOps all at once leads to context collapse, no accountability, and inconsistent quality.
The solution: A structured workspace where each agent plays a defined role in a gated pipeline:
PLAN → BUILD → TEST → VERIFY → SHIP
- Orchestrator coordinates — never writes code itself
- Architects design before developers build
- Reviewers validate independently after each step
- Safety hooks prevent dangerous commands, enforce commit conventions, auto-lint
Each step has an approval gate. Failures retry (3×) → helper role (2×) → human escalation. The human only sees escalations, not every step.
| Scenario | multiagent-template | Single Claude Code | Standalone Aider |
|---|---|---|---|
| Pair-programming a single file | — | ✅ | ✅ |
| Full feature from spec to merged PR | ✅ | Struggles | Struggles |
| Team of specialized roles | ✅ | ❌ | ❌ |
| Autonomous overnight backlog drain | ✅ | Limited | ❌ |
| Works inside your existing IDE | ✅ (via provider flags) | ✅ | ✅ |
| Any AI provider (Claude, Gemini, Codex…) | ✅ 13 providers | Claude only | Any, but no pipeline |
| Provider | Binary / Tool | Best for | Notes |
|---|---|---|---|
| claude | claude | Terminal-first, most capable | Claude Code by Anthropic — default |
| nessy | nessy | Claude users with a different alias | Nessy CLI — Claude-compatible |
| codex | codex | OpenAI API key holders | OpenAI Codex CLI — creates AGENTS.md |
| qwen | qwen-code | Open-source model users | Qwen Code — creates QWEN.md |
| gemini | gemini | Google API key holders | Gemini CLI — creates GEMINI.md |
| cursor | Cursor IDE | IDE-first workflows | Rules in .cursor/rules/ (MDC format) |
| windsurf | Windsurf IDE | IDE-first + Codeium users | Rules in .windsurf/rules/ (Wave 8+) |
| copilot | GitHub Copilot | VS Code + GitHub users | Reads .github/copilot-instructions.md |
| cline | Cline | VS Code extension users | Creates .clinerules in project root |
| aider | Aider | Terminal pair programming | Creates .aider.conf.yml — auto-reads CLAUDE.md |
| continue | Continue.dev | VS Code + JetBrains users | .continue/config.yaml with /orchestrator slash command |
| roo | Roo Code | VS Code extension users | Rules in .roo/rules/ — auto-loaded per project |
| kiro | Amazon Kiro | AWS + VS Code users | Steering docs in .kiro/steering/ — auto-loaded by Kiro |
Use --provider all to scaffold all providers (claude + nessy + codex + qwen + cursor + windsurf + copilot + gemini + cline + aider + continue + roo + kiro).
One human (CEO) gives tasks. The Orchestrator agent breaks them into steps, picks the right specialist role, runs the pipeline, and delivers a PR. Human escalation is required only for: public content, breaking API changes, infra decisions with cost impact, or 5+ consecutive failures.
| Type | Steps | When |
|---|---|---|
feature | PLAN → BUILD → TEST → VERIFY → SHIP | New functionality |
bugfix | BUILD → TEST → VERIFY → SHIP | Skip planning |
infra | PLAN → BUILD → VERIFY → SHIP | No test step |
content | PLAN → BUILD → VERIFY(CEO) | Docs / marketing |
spike | PLAN | Research only |
| Mode | How to trigger | Description |
|---|---|---|
| CEO Mode | /orchestrator <task> | Human gives task, orchestrator executes |
| Single Expert | /<role> <question> | Direct expert call, no pipeline |
| Autonomous | claude -p "/orchestrator ..." | Orchestrator self-selects tasks from backlog |
After multiagent-setup new MyProject --provider cursor (or any IDE provider):
Open the workspace in your IDE — open the
MyProject/directory (not a subdirectory)Invoke the orchestrator — use your IDE's AI chat panel and type
/orchestrator <task>:- Cursor:
Cmd+KorCmd+Lchat panel →/orchestrator Implement user authentication - Windsurf: Cascade panel →
/orchestrator ... - Copilot: GitHub Copilot Chat →
@workspace /orchestrator ... - Cline: Cline panel →
/orchestrator ... - Continue / Roo / Kiro: respective chat panel →
/orchestrator ...
- Cursor:
Slash commands are pre-loaded —
.cursor/rules/orchestrator.mdc(and equivalent for each provider) is scaffolded withalwaysApply: true, so the orchestrator role is active automatically. All 20+ specialist roles are in.claude/commands/and invokable as slash commands.Hooks require the CLI on PATH — safety hooks (
block-dangerous,auto-lint, etc.) are compiled into themultiagent-setupbinary. IDE agents don't run hooks unless the binary is on PATH and the IDE inherits the shell environment. Terminal providers (Claude, Codex, Gemini, etc.) get hooks automatically via.claude/settings.json.The pipeline is the same —
PLAN → BUILD → TEST → VERIFY → SHIPworks identically whether you're using a terminal agent or an IDE agent. Gate approvals appear as responses in the chat panel.
To run the orchestrator fully headlessly (no human in the loop):
1. Create a GitHub Project
Go to your GitHub repo → Projects → New project. Use the "Board" template with columns: Backlog, In Progress, Done. Link the project to your repository.
2. Add issues to Backlog
Write issues as tasks for the orchestrator. Good format:
Title: Implement user authentication with JWT
Body: - Users can sign up with email/password
- JWT tokens expire in 7 days
- Refresh token flow included
- Acceptance: all auth endpoints have integration tests
3. Trigger autonomously
# One-shot: pick one task from backlog and implement it
claude -p "/orchestrator Pick the highest-priority task from the GitHub Project backlog and implement it."# Or use the GitHub Actions CI workflow (label an issue with 'orchestrator')# The scaffolded .github/workflows/orchestrator.yml triggers automaticallyThe orchestrator.yml CI workflow triggers when you add the orchestrator label to any issue. It runs headlessly, implements the task, and opens a PR. You review and merge.
4. Loop mode (drain the whole backlog unattended)
claude -p "/orchestrator Pick tasks from the GitHub Project backlog. Implement each one. Create a PR per task. Stop when the backlog is empty."The
CLAUDE.mdtemplate variables{{GITHUB_ORG}}and{{GITHUB_REPO}}are substituted at workspace creation time with your actual org/repo names — no manual editing needed.
MyProject/
├── CLAUDE.md <- workspace context (read by AI on every session)
├── code/ <- product repo (git-ignored)
├── docs/
│ ├── process.md <- operational manual (pipeline source of truth)
│ ├── role-capabilities.md <- role index for dynamic orchestrator routing
│ └── workflows/ <- pipeline specs (WORKFLOW-*.md)
├── .claude/
│ ├── commands/ <- slash-command roles (synced from agency-agents)
│ ├── hooks/lint.json <- auto-lint formatter config
│ ├── mcp.json <- MCP server config
│ └── settings.json <- hook configuration
├── .github/
│ └── workflows/
│ └── orchestrator.yml <- run orchestrator in CI (claude/nessy providers)
├── .codex/ <- Codex config (--provider codex)
│ └── skills/ <- orchestrator skill pre-loaded
└── tools/
├── completions.zsh <- zsh completions
└── completions.ps1 <- PowerShell completions
The workspace uses two separate instruction files:
| File | Purpose |
|---|---|
CLAUDE.md (workspace root) | How Claude operates — pipeline, roles, process, team structure |
code/<project>/CLAUDE.md | What Claude builds — architecture, patterns, build commands, test setup |
When you clone your code repo into code/<project>/, create a CLAUDE.md there describing the codebase (or copy your existing one). Claude Code automatically reads both files.
Bringing in an existing codebase:
multiagent-setup new MyProject
cd MyProject
git clone https://github.com/my-org/my-repo code/MyProject
# Edit CLAUDE.md to describe the project → doneAll hooks are compiled into the multiagent-setup binary — no shell scripts, no platform quirks.
| Hook | Trigger | Action |
|---|---|---|
block-dangerous | PreToolUse (Bash) | Blocks rm -rf /, push --force main, DROP TABLE, etc. |
enforce-commit-msg | PreToolUse (Bash) | Enforces conventional commits (feat:, fix:, etc.) |
auto-lint | PostToolUse (Edit/Write) | Runs formatter on changed file (prettier, ruff, gofmt, rustfmt…) |
log-agent | PreToolUse (Agent) | Logs sub-agent launches to .claude/agent-log.jsonl |
stop-guard | Stop | Reminds to run tests and update knowledge graph |
research-reminder | PostToolUse (WebSearch) | Reminds to persist research in O'Brien memory |
Built on agency-agents — a community-maintained library of specialist Claude Code slash commands.
multiagent-setup sync-rolesclones and installs the full set at workspace creation. Go star that repo too — it's what makes the roles work.
20+ specialist roles installed automatically:
| Layer | Roles |
|---|---|
| Strategy | /product-manager, /product-trend-researcher |
| Management | /orchestrator, /testing-reality-checker, /specialized-workflow-architect |
| Engineering | /engineering-software-architect, /engineering-backend-architect, /engineering-frontend-developer, /engineering-code-reviewer, /engineering-devops-automator, /engineering-security-engineer |
| AI / ML | /engineering-ai-engineer |
| Design | /design-ux-researcher, /design-ui-designer |
| GTM | /specialized-developer-advocate, /engineering-technical-writer, /marketing-content-creator |
The orchestrator routes dynamically via docs/role-capabilities.md — no hardcoded assignments. If no role fits, it creates an ad-hoc role on the fly.
Graph knowledge base on PostgreSQL + Apache AGE, connected via age-mcp. Stores modules, pipelines, role bindings, security findings, code insights. Grows with every task.
Semantic memory on pgvector — cross-session context, task locking, crash recovery.
multiagent-setup install-mcps # interactive Docker setup
multiagent-setup install-mcps --manual # enter connection strings manuallySee examples/ for concrete workflows:
- SaaS Starter — foundation, auth, billing, autonomous sessions
- Open Source Maintainer — bug triage, PR reviews, releases
multiagent-setup new <project> [org] [--provider claude|nessy|codex|qwen|cursor|windsurf|copilot|gemini|cline|aider|continue|roo|kiro|all] [--template default|saas|oss|internal]
multiagent-setup init [dir] [--provider <name>] [--template <name>] [--force] # add workspace to existing repo
multiagent-setup add-provider <provider> [--force] # add provider to existing workspace
multiagent-setup remove-provider <provider> [--force] [--dry-run] # remove a provider (--dry-run to preview)
multiagent-setup list-providers # list installed and available providers
multiagent-setup update [--force] [--dry-run] # update templates (--dry-run to preview)
multiagent-setup sync-roles [--clone|--pull] [--global] [--agency-dir <path>]
multiagent-setup install-mcps [--docker|--manual] [--age-conn <str>] [--obrien-conn <str>] [--target <dir>]
multiagent-setup hook <name>
multiagent-setup doctor [--for sync-roles|init|update] # check workspace or run pre-flight check
multiagent-setup completions zsh|pwsh # print shell completion script
multiagent-setup -v | --version| Tool | macOS/Linux | Windows |
|---|---|---|
| .NET SDK 10+ | brew install dotnet | winget install Microsoft.DotNet.SDK.10 |
| GitHub CLI | brew install gh | winget install GitHub.cli |
| git, jq | brew / apt | winget install Git.Git jqlang.jq |
| Agent CLI | see provider table above | same |
| Docker | optional, for AGE/O'Brien | winget install Docker.DockerDesktop |
bootstrap.sh / bootstrap.ps1 install everything automatically on a clean machine.
Templates live in tools/setup-cli/Templates/. Each provider gets its own directory under providers/. See CONTRIBUTING.md for setup instructions and how to add a new provider.
Questions and ideas → Discussions.
If this saves you time, a ⭐ star helps others find it.
Does this work with projects that already have code?
Yes — two options: (1) run multiagent-setup init . inside your existing repo to add workspace files directly, or (2) run multiagent-setup new MyProject to create a separate workspace, then clone your repo into code/MyProject/.
How do I use multiple providers?
Run multiagent-setup new MyProject --provider all to scaffold all providers at once. To add a provider to an existing workspace: multiagent-setup add-provider gemini.
What does the orchestrator do when I'm not watching?
In CEO Mode it waits for your next task. In Autonomous mode (claude -p), it picks tasks from the GitHub Project backlog and escalates only for defined edge cases.
Can I add custom roles?
Yes. Create a .md file in .claude/commands/ with a name: frontmatter field. The orchestrator uses it automatically, and can also create ad-hoc roles on the fly.
Do I need Docker / AGE graph / O'Brien?
No. The MCPs are entirely optional. The workspace and pipeline work without them. install-mcps is a separate optional step for teams that want persistent memory across sessions.
A safety hook is blocking a command I need.
Claude Code will prompt you to approve or deny the hook's decision. Click "Allow" to override for that command. You can also edit .claude/settings.json to adjust which hooks are active.
How do I update an existing workspace to the latest templates?
Run multiagent-setup update from inside your workspace. Your customised context files (CLAUDE.md, GEMINI.md, etc.) are preserved. Add --force to overwrite everything.
Where can I see what agents were spawned in a session?
The log-agent hook writes to .claude/agent-log.jsonl — one JSON line per agent launch with timestamp, type, model, and prompt preview.
Is this production-ready?
Yes. The tool has a stable API (breaking changes only in major version bumps). It's used for autonomous overnight runs and SaaS development sessions. The doctor command checks your workspace for common config issues.
Can the orchestrator run in CI without a local machine?
Yes. A GitHub Actions workflow (orchestrator.yml) is scaffolded for claude/nessy workspaces. It triggers on issue labels and runs the orchestrator headlessly via claude -p. See examples/ for autonomous session patterns.
Does the pipeline need internet access?
Only for git push/PR creation steps and any web research the agent does. All local code analysis, formatting, and linting runs offline.