For the past few months, I've been building CrewForm — an open-source platform for creating and orchestrating multi-agent AI workflows through a web UI.
Think of it as the visual, no-code alternative to frameworks like CrewAI or LangGraph. Instead of writing Python scripts to wire agents together, you configure everything in a dashboard.
I wanted to share what it does, how it's built, and some lessons from the journey.
What it does
You create AI agents by picking a model (15 LLM providers supported — OpenAI, Anthropic, Gemini, Groq, Ollama, etc.), writing a system prompt, and attaching tools.
Then you combine agents into teams that run in one of three modes:
Pipeline — agents execute sequentially, each passing output to the next. Great for content workflows (research → write → edit → format).
Orchestrator — a "brain" agent receives the task, breaks it down, and delegates sub-tasks to specialist workers. Best for complex, dynamic problems.
Collaboration — agents discuss in a shared thread, debate approaches, and converge on a consensus. Useful for analysis and decision-making.
Each mode has a visual workflow canvas where you can see the execution graph — pipeline chains, orchestrator delegation trees, or collaboration circles.
The protocol story
One thing I'm particularly excited about is protocol support. CrewForm implements all three emerging agentic protocols:
MCP (Model Context Protocol) — your agents can connect to any MCP-compatible tool server. There are thousands of these now — file systems, databases, APIs, web scrapers. One config, and your agent has new capabilities.
A2A (Agent-to-Agent) — agents in CrewForm can collaborate with agents from completely different frameworks. Publish your agent as an A2A endpoint, or consume external ones.
AG-UI (Agent-User Interaction) — real-time SSE streaming so you can build frontends that show what agents are doing as they work.
Self-hosting & local models
This was a non-negotiable. If you're processing sensitive data through AI, you need to own the infrastructure.
bash
docker compose up -d
That's the entire setup. Pair with Ollama, and you get fully local AI — Llama 3.3, DeepSeek R1, Qwen 2.5, Mistral, and more. Zero API keys, zero data leaving your network.
Knowledge Base (RAG)
Upload documents (TXT, MD, CSV, JSON), CrewForm chunks, and embeds them using pgvector, and agents search the knowledge base at runtime. No external vector DB needed — it's all in Postgres.
Tech stack
For anyone curious about the architecture:
Frontend: React + TypeScript + Vite, TailwindCSS, React Flow (workflow canvas)
Backend: Supabase — Auth, PostgreSQL with RLS, Realtime, Edge Functions, Storage
Task runner: Node.js — polls for pending tasks, runs the execution engine
Vectors: pgvector extension in Postgres
Hosting: Vercel (frontend), Railway (task runner), Supabase Cloud (backend)
Self-hosted: Docker Compose for the full stack
The decision to build on Supabase was pivotal. Row Level Security means every database query is automatically scoped to the user's workspace — security by default, not by convention. And real-time subscriptions meant live task execution updates without managing WebSocket servers.
Bring Your Own Keys
CrewForm is BYOK — bring your own API keys. You plug in your keys for whatever providers you use, and we never touch them. No per-token markup, no middleman.
Currently supporting: OpenAI, Anthropic, Google Gemini, Groq, Mistral, Cohere, Perplexity, DeepSeek, Together AI, Fireworks, OpenRouter, xAI, Cerebras, HuggingFace, and Ollama.
What's next
Interactive workflow canvas (drag-to-add agents, connect nodes to define execution order)
Cost optimisation dashboard
Enterprise features (SSO, audit log streaming)
More marketplace templates
Try it
🔗 GitHub: github.com/CrewForm/crewform
🌐 Website: crewform.tech
📚 Docs: docs.crewform.tech
💬 Discord: discord.gg/TAFasJCTWs
Licensed under AGPL-3.0. Would love feedback from the dev community — especially on the self-hosting experience and what features you'd want to see next!
Top comments (0)