The open-source visual studio for CrewAI
Build multi-agent workflows in the browser · Run with live observability · Export clean Python · No lock-in
Quick Start · Features · Architecture · Docs · Contributing
CrewForge is a self-hosted, no-code studio that puts a visual front door on CrewAI. Design a multi-agent crew in the browser, run it live, watch events stream in real-time, then export a clean, idiomatic Python project you own completely.
No API key required to start — dry-run mode uses a built-in mock LLM so you can build and explore without touching a provider.
| 🎨 Visual workflow builder | Drag-and-drop canvas with agent/task nodes, live run glow, and a full inspector |
| 🤖 103 built-in tools | Searchable catalog; configure API keys per tool and attach to any agent |
| 🔌 MCP integrations | Connect local or remote MCP servers; browse the official registry with security ratings |
| 📚 Knowledge bases | Keyless vector RAG (fastembed) + Kuzu knowledge graph; ingest text, files, URLs, GitHub repos |
| 🏃 Batch runs | Paste a CSV → one tracked run per row, with aggregate cost and per-row drill-in |
| 📅 Scheduling + webhooks | Cron schedules and webhook triggers with per-run input pinning |
| 🧪 Eval suite | Define test cases, score with contains/regex/equals or LLM-as-judge |
| 💰 Cost tracking | Per-run estimated spend from a curated pricing table; shown on the timeline and dashboard |
| 🛑 Human-in-the-loop | Approve, edit output, or request changes mid-run from the Runs page |
| 📤 Code export | Download a complete, runnable CrewAI Python project — no vendor lock-in |
| 🔑 Secrets vault | Provider keys encrypted at rest (Fernet); never returned to the client |
| 🌐 Multi-LLM | Per-workflow and per-agent LLM selection; any OpenAI-compatible provider |
| 🔁 Replay & duplicate | Re-run any past run with the same inputs; duplicate workflows in one click |
| ⚡ Dry-run mode | Built-in mock LLM; zero network calls; every feature works without a key |
Prerequisites: Python 3.10+, uv, Node.js 18+
git clone https://github.com/Arturski/CrewForge.git
cd CrewForge
uv sync
npm --prefix web install && npm --prefix web run build
uv run crewforge
# → open http://localhost:8765No API key needed. The demo workspace runs immediately in dry-run mode.
- Settings → Models → Add connection
- Choose a provider preset or enter a custom base URL
- Paste your API key → Test connection
- In the Builder, toggle Dry run off → Run
Works with any OpenAI-compatible provider: OpenAI, Anthropic, MiniMax (api.minimaxi.chat/v1), Ollama, Groq, Together AI, and hundreds more via LiteLLM.
# Terminal 1 — API with hot reload
uv run uvicorn server.app:app --reload --port 8765
# Terminal 2 — Vite dev server
npm --prefix web run dev # → http://localhost:5180
Key design decisions:
- Single-process, SQLite-backed — zero external dependencies; runs on a laptop or a tiny VPS
- Compatibility engine —
compiler/manifest.pyintrospects the installedcrewaiat startup, building a field manifest that drives the UI dynamically. New CrewAI fields surface as form controls with no code changes. - Dry-run everywhere —
FakeLLMimplementsBaseLLMand returns a ReActFinal Answerso agents complete cleanly. Every feature is exercisable without a key. - No lock-in — the exporter emits a clean, standalone CrewAI project. CrewForge is a tool to build and operate, not a runtime you depend on.
| Doc | Description |
|---|---|
| Quick Start Guide | Step-by-step: first workflow, going live, batch & eval |
| Architecture | Module breakdown, data model, design decisions |
| Deployment | Self-hosting, reverse proxy, Docker, environment variables |
| API Reference | All REST endpoints with request/response shapes |
| Contributing | Dev setup, code conventions, adding new features |
| Roadmap | What's shipped and what's next |
| Security | Threat model, key storage, hardening roadmap |
| Layer | Technology |
|---|---|
| Backend | Python 3.10+, FastAPI, SQLite (WAL), uvicorn |
| AI framework | CrewAI 1.14+, crewai-tools |
| Embeddings | fastembed (bge-small-en-v1.5, keyless, local) |
| Knowledge graph | Kuzu (embedded, per-KB) |
| Encryption | cryptography (Fernet) |
| Scheduling | croniter |
| Frontend | React 19, Vite, TypeScript, Tailwind v4 |
| UI components | Radix UI primitives + shadcn/ui-style system |
| Canvas | XyFlow (React Flow) |
| Icons | lucide-react |
CrewForge/
├── server/ # FastAPI backend
│ ├── app.py # All REST + SSE endpoints (~50 routes)
│ ├── store.py # SQLite: all CRUD (workspaces, runs, KBs, …)
│ ├── runner.py # Run execution, event capture, HITL, cancel
│ ├── batches.py # Batch-run driver (CSV → N tracked runs)
│ ├── evals.py # Eval suite + LLM-as-judge scoring
│ ├── schedules.py # Cron daemon + webhook trigger handler
│ ├── llms.py # Multi-LLM connection management
│ ├── mcp.py # MCP server connect/discover/registry
│ ├── secrets.py # Fernet encryption at rest
│ ├── pricing.py # Token cost estimates
│ ├── compiler/
│ │ ├── manifest.py # CrewAI field introspection → UI manifest
│ │ ├── adapter.py # Spec → live Crew/Agent/Task + FakeLLM
│ │ ├── exporter.py # Spec → runnable CrewAI project (zip)
│ │ └── tools.py # 103 built-in crewai_tools catalog
│ └── knowledge/
│ ├── kb.py # KB CRUD + background ingest orchestration
│ ├── vector.py # fastembed chunking + cosine search
│ ├── graph.py # Kuzu graph: entities, relations, hybrid retrieval
│ ├── extract.py # LLM entity/relation extraction
│ ├── web.py # URL ingest + same-host crawl
│ └── github.py # GitHub repo ingest via tarball
├── web/ # React 19 SPA
│ └── src/
│ ├── pages/ # Dashboard, Builder, Runs, Tools, Knowledge, Settings, Code
│ └── components/ # CrewCanvas, EventTimeline, ui.tsx, Sidebar
├── tests/ # 52 pytest tests (no network required)
├── docs/ # This documentation + architecture diagrams
└── spikes/ # Architecture proof-of-concept scripts
# 52 tests, ~5 seconds, no network, no API key
uv run --extra dev pytest -q
# Lint
uv run --extra dev ruff check server tests
# TypeScript strict + production build
npm --prefix web run buildAll tests pass green. Zero browser console errors (verified with Playwright across all pages + mobile viewport).
Contributions are very welcome — see CONTRIBUTING.md for the dev setup, code conventions, and how to add support for new CrewAI features.
git clone https://github.com/Arturski/CrewForge.git && cd CrewForge
uv sync && npm --prefix web install
uv run --extra dev pytest -q # all green
npm --prefix web run build # tsc strict passSee ROADMAP.md for what's shipped and what's coming. Highlights on deck:
- Eval history & trend — pass-rate over time per workflow
- CrewAI Flows — visual multi-crew orchestration with
@router/@listen - Run analytics — cost, token, and success-rate dashboards
- Workspace versioning — snapshot, diff, rollback
- Containerized workers — durable, crash-safe execution
Apache-2.0 — free to use, modify, and distribute.