What it does

Giselle is a browser-based visual builder for AI agent workflows. You drag, drop, and connect nodes to compose multi-step pipelines that can call OpenAI, Anthropic, or Google models. The project is a Next.js app you run locally with pnpm turbo dev after dropping an API key into .env.local.

The interesting bit

The “multi-model composition” angle is the hook: instead of locking you to one provider, each node in a workflow can pick its own LLM. There’s also GitHub vector-store integration for RAG, though the README is light on how that plumbing actually works.

Key highlights

• Visual drag-and-drop agent builder (no code required)
• Mix-and-match models: GPT, Claude, Gemini in the same workflow
• Self-hostable or use the managed cloud tier (30 min free agent time/month)
• GitHub-flavored RAG via a “Knowledge Store”
• “Vibe Coding Guide” for non-engineers using AI assistants to modify the project

Caveats

• Team collaboration and template hub are marked “In Development”
• Public roadmap is still being drafted; README says it will be “updated accordingly”
• Only 533 stars and a thin contributor guide—expect rough edges

Verdict

Worth a spin if you want a self-hosted alternative to commercial agent platforms like LangFlow or Flowise. Skip it if you need mature multi-user features or detailed docs today; the project is still finding its footing.