Visual LangGraph builder where your graph runs exactly as drawn — visually compose nodes, edges, and tools, then stream execution in real time.
Note
Graphweave is in early development. The canvas builder and execution backend are functional, but graph execution (run a graph and stream results) is not yet wired end-to-end.
Every team whiteboard-designs their agent flow, then spends days translating it into LangGraph code — wiring nodes, debugging edges, mapping state. The diagram and the code drift apart immediately.
Graphweave kills that gap. The canvas is the execution. No translation step, no drift. You draw the agent, you run the agent. Same artifact.
What you draw is what runs. GraphSchema maps 1-to-1 to LangGraph primitives. No abstraction layer between the visual and the execution.
| Layer | What works | What's next |
|---|---|---|
| Canvas | Home view, Start/LLM/End nodes, edge wiring, config panel, save/load, floating toolbar | SSE run panel, full node set → |
| Execution | API key auth, graph CRUD, tool registry (8 tools), migrations | Graph run + SSE streaming → |
| Shared | GraphSchema contract, node types, edge types | Condition + HumanInput wiring |
packages/
├── canvas/ React 19 + React Flow + Zustand — the visual builder
├── shared/ GraphSchema types — the contract between canvas and execution
├── sdk-core/ Plugin interfaces (transport, events, storage)
└── execution/ FastAPI + LangGraph — builds and runs graphs via SSE
The canvas produces a GraphSchema. The execution layer consumes it, builds a LangGraph StateGraph, and streams node-by-node execution back over SSE.
| Node | Purpose |
|---|---|
| Start | Entry point — every graph has exactly one |
| LLM | Calls a language model (Gemini, OpenAI, Anthropic) |
| Tool | Runs a registered tool with input/output mapping |
| Condition | Branches on field checks, LLM routing, tool errors, or iteration limits |
| Human Input | Pauses execution and waits for user input |
| End | Terminal node — every graph has at least one |
- Google Gemini
- OpenAI
- Anthropic
- Node.js >= 20
- pnpm >= 10
- Docker (for the execution layer)
- uv (for Python dependency management)
# Clone
git clone https://github.com/prosdevlab/graphweave.git
cd graphweave
# Install dependencies
pnpm install
# Copy environment config
cp .env.example packages/execution/.env
# Add at least one LLM provider key to packages/execution/.env# Start everything (canvas + execution in Docker)
pnpm devThis runs the React canvas on http://localhost:5173 and the FastAPI execution server on http://localhost:8000.
# TypeScript + Python checks in parallel
pnpm verify
# Or individually
pnpm typecheck # TypeScript type checking
pnpm check # Biome lint + format
pnpm verify:py # Ruff + pytest| Path | What | Managed by |
|---|---|---|
packages/canvas/ | React 19 + Vite + Tailwind v4 | pnpm |
packages/shared/ | GraphSchema TypeScript types | pnpm + tsup |
packages/sdk-core/ | Plugin interfaces | pnpm + tsup |
packages/execution/ | FastAPI + LangGraph | uv + Docker |
docker-compose.yml | Production execution | Docker |
docker-compose.dev.yml | Dev execution (hot reload) | Docker |
This repo uses Conventional Commits enforced by commitlint.
type(scope): description
# Examples
feat(canvas): add node drag-and-drop
fix(execution): handle missing API key gracefully
chore(deps): bump langchain to 0.3
Types:feat, fix, docs, style, refactor, perf, test, chore, ci, build, revert
Scopes:canvas, shared, sdk-core, execution, docs, deps, docker, schema, skills
Commits automatically run:
.envfile guard — blocks accidental secret commits- lint-staged — Biome for TS/JS, Ruff for Python
- typecheck — full
tsc --noEmit(cached by Turbo) - commitlint — validates commit message format
MIT