I'm a photographer. I built a DSL for multi-agent workflows

A photographer has built AgentFlow, a declarative DSL that lets users define multi-agent AI workflows in simple `.aflow` files and compile them into MCP tools for Claude Code without writing any integration code. The system supports assigning different AI models to different agents, running iterative loops with structured output, and includes a CLI with validation, execution, and MCP configuration commands. This allows non-programmers to create complex, multi-model AI pipelines that are git-friendly and reviewable by anyone.

A declarative language for multi-agent AI workflows — compile to MCP tools, no code required. Write complex multi-agent workflows in clean, readable .aflow files. Each workflow becomes a tool in Claude Code via MCP — zero integration code. Assign different AI models to different agents, run iterative loops, and let agents collaborate with structured output. workflow code quality description: "Iterative code review with writer, tester, and critic" version: "1.0.0" agents: agent writer → model: "local-fast" agent tester → model: "openrouter-smart" agent critic → model: "claude-sonnet" loop quality gate phases: write, test, review repeat while: review.verdict == "needs work" max iterations: 5 done when: review.confidence = 0.85 | Traditional LangGraph, CrewAI | AgentFlow DSL | | |---|---|---| Define workflows | Python code ~40 lines boilerplate | .aflow file ~20 lines | Multi-model | Manual provider switching | Per-agent model aliases | MCP integration | Write MCP server code | Automatic — each workflow = MCP tool | Git-friendly | Code + config scattered | Single .aflow file | Reviewable | Need Python knowledge | Readable by anyone | npm install -g @anhonestboy/agentflow agentflow init Interactive wizard: choose providers Claude, OpenRouter, Ollama , configure model aliases, save API keys. Create my-workflow.aflow : workflow blog post description: "Generate and refine a blog post" version: "1.0.0" agents: agent researcher mode: patient must produce: - outline - key points agent writer mode: focused must produce: - draft - word count: int agent editor mode: adversarial must produce: - verdict - suggestions - confidence: float phases: phase research agent: researcher input: trigger.topic output: outline, key points phase write agent: writer input: research.outline, research.key points output: draft, word count phase edit agent: editor input: write.draft output: verdict, suggestions, confidence loop revision cycle phases: write, edit repeat while: edit.verdict == "needs work" max iterations: 3 on each iteration: send to: writer payload: edit.suggestions done when: edit.confidence = 0.8 and edit.verdict == "approved" agentflow check my-workflow.aflow Validate agentflow run my-workflow.aflow --input 'topic="AI in photography"' agentflow mcp-config Copy the JSON output to your Claude Code MCP settings. Your workflow is now a tool — call it directly from Claude Code. | Provider | Status | Notes | |---|---|---| Claude Anthropic | ✅ | Native SDK, multi-round tool use | OpenRouter | ✅ | 315+ models, automatic provider routing | Ollama | ✅ | Local execution, no API key needed | Configure model aliases for cost optimization — use cheap models for drafting, frontier models for review: { "models": { "local-fast": { "provider": "ollama", "model": "qwen3:8b" }, "openrouter-smart": { "provider": "openrouter", "model": "google/gemini-2.5-flash" }, "claude-sonnet": { "provider": "claude", "model": "claude-sonnet-4-5" } } } agentflow init Interactive setup wizard agentflow check <file Validate workflow + summary agentflow run <file --input '…' Execute with real LLMs agentflow run <file --mock Execute with mock agents no API key needed agentflow compile <file Compile to IR JSON agentflow validate <file Validate only no summary agentflow mcp-config Print MCP server config for Claude Code agentflow models List configured models + connectivity agentflow resume <file --instance <uuid Resume interrupted workflow agent <id model: "<alias " Model alias from config default: "auto" mode: <mode focused | adversarial | reliable | precise | strict | patient | objective tools: <name , ... Built-in tools: file write, file read, shell exec, test runner must produce: - <name Required output field string - <name : float Typed output field constraint: "<rule " Natural language constraint phase <id agent: <agent id input: <ref , ... trigger.field or phase id.output output: <name , ... inject context: "<path " Optional: inject file content into agent context timeout: 30min For human action required phases loop <id phases: <phase id , ... repeat while: <condition review.verdict == "needs work" max iterations: <n on each iteration: send to: <agent id payload: <ref Feedback to inject on max exceeded: escalate to: <agent id message: "<... " Comparison review.confidence = 0.85 Logical review.verdict == "approved" and review.confidence = 0.85 not review.verdict == "needs work" .aflow file │ ▼ Tokenizer ──► Parser ──► Compiler AST → IR │ ┌─────────▼──────────┐ │ Validator S1-S10 │ └─────────┬──────────┘ │ ┌─────────▼──────────┐ │ WorkflowRunner │ │ ┌───────────────┐ │ │ │ ExecutorResolver│ │ │ │ ┌─────────────┐│ │ │ │ │ Claude ││ │ │ │ │ OpenRouter ││ │ │ │ │ Ollama ││ │ │ │ └─────────────┘│ │ │ └───────────────┘ │ └─────────┬──────────┘ │ ┌─────────▼──────────┐ │ MCP Server │ │ stdio JSON-RPC │ └─────────┬──────────┘ │ Claude Code / Cursor | File | Description | |---|---| examples/blog-post.aflow | Researcher → writer → editor with revision loop | examples/code-quality.aflow | Writer → tester → critic with quality gate loop | examples/code-quality-with-plan.aflow | Extended with planning phase | examples/custom-domain.aflow | 7-phase domain provisioning workflow | git clone https://github.com/anhonestboy/agentflow.git cd agentflow npm install npm run build npm test 92 tests, 6 suites npm run dev -- check examples/code-quality.aflow v1.1 — VS Code extension syntax highlighting, LSP v1.2 — Parallel phase execution v1.3 — Workflow registry & sharing v2.0 — Web visualizer, CI/CD integration MIT — see LICENSE /anhonestboy/agentflow/blob/main/LICENSE for details.