CrewForm
A self-hostable runtime for orchestrating multiple AI agents that natively supports the MCP, A2A and AG-UI protocols, letting agents call outside tools, interoperate with agents built on other frameworks, and stream their state to a frontend in real time.
CrewForm lets a team build, run and observe multi-agent systems across tools, agents, models, memory, workflows and frontends without depending on one vendor's stack. As an MCP client it lets agents discover and call tools from any MCP server, and as an MCP server it can expose its own agents as tools to clients such as Claude Desktop or Cursor; as an A2A participant it can delegate tasks to external agents or publish its own agents for other systems to call, and AG-UI streams text, tool calls and state transitions to a frontend over SSE, including approval and confirmation interactions.
A visual, drag-and-drop workflow canvas supports three team modes, pipeline, orchestrator and collaboration, with live execution visualization, and a knowledge base combines pgvector, full-text search and reranking for retrieval-augmented generation. It supports 16 LLM providers, including OpenAI, Anthropic, Gemini, Groq and Ollama for fully local, air-gapped operation.
It is written in TypeScript, self-hosted on Supabase, and released under the AGPL-3.0 license.