The examples/ folder holds small apps that show how to use NodeTool from your own code. Each one builds on the monorepo packages, which are not published to npm, so build them from the repository root first:
nvm use
npm install
npm run build:packages
| Example | Shows | Runs against |
|---|---|---|
chat_app |
A React chat client built on the SDK | A running NodeTool server |
nextjs-vercel |
Workflows run inside a Next.js route on Vercel | No separate server |
nextjs-cloudflare |
The same app on Cloudflare Workers | No separate server |
workflow_runner |
A static page that runs workflows over the WebSocket API | A running NodeTool server |
chat_app
A chat client for the NodeTool server, built with Vite, React 19, Tailwind and vendored shadcn/ui components. It shows how to use the typed tRPC client and the streaming chat WebSocket from createNodetoolClient(): listing threads and messages, picking a language model, and streaming assistant replies.
Packages: @nodetool-ai/sdk.
Run it. Start the server with the fake provider enabled, then start the example in a second terminal:
NODETOOL_ENABLE_FAKE_PROVIDER=1 npm run dev:nodetool -- serve --port 7777
npm run dev --workspace=@nodetool-ai/example-chat-app
Open http://localhost:5173. The Vite dev server proxies /trpc, /api and /ws to http://localhost:7777. Set PROXY_API_TARGET if the server runs elsewhere.
Smoke test. With the server running, this drives the chat flow without a browser and prints PASS or FAIL for each stage:
npm run live-test --workspace=@nodetool-ai/example-chat-app
nextjs-vercel
A Next.js App Router app that runs NodeTool workflows on the server and streams the result to the browser as Server-Sent Events. The route wraps a NodeRegistry in createWorkflowHandler from @nodetool-ai/workflow-runner, so no separate backend runs. The registry imports node groups from per-file subpaths such as @nodetool-ai/core-nodes/nodes/*, because the package index pulls in native bindings that cannot deploy serverless. An optional OpenAI Web Search sample reads OPENAI_API_KEY from the server environment, so the key never reaches the browser.
Packages: @nodetool-ai/workflow-runner, @nodetool-ai/node-sdk, @nodetool-ai/runtime, @nodetool-ai/protocol, @nodetool-ai/core-nodes, @nodetool-ai/llm-nodes.
Run it:
cd examples/nextjs-vercel
npm run dev
Open http://localhost:3002. For the LLM sample, put OPENAI_API_KEY=sk-... in .env.local first.
Deploy to Vercel:
npm i -g vercel
vercel
vercel --prod
Other scripts in package.json: build, start (port 3002), lint and typecheck.
nextjs-cloudflare
The same app as nextjs-vercel, deployed to Cloudflare Workers through OpenNext. The runner uses platform: "workers" and registry.forPlatform("workers") keeps only Workers-compatible nodes. wrangler.jsonc enables nodejs_compat, and OpenNext bundles the @nodetool-ai/* packages into the Worker. Nodes that need native modules or child_process (image, video, audio, vector, and code runners) are excluded.
Packages: the same six as nextjs-vercel, plus @opennextjs/cloudflare and wrangler as dev dependencies.
Run it:
cd examples/nextjs-cloudflare
npm run dev # Next dev server on http://localhost:3003
npm run preview # build and run on a local workerd
For the LLM sample, put OPENAI_API_KEY=sk-... in .dev.vars before npm run preview.
Deploy to Cloudflare:
npx wrangler login
npm run deploy
Set the key on the deployed Worker with npx wrangler secret put OPENAI_API_KEY. The cf-typegen script regenerates the Cloudflare environment types.
workflow_runner
A static HTML and JavaScript page that connects to the NodeTool server over the WebSocket API, loads a workflow, runs it, and shows progress, outputs and a graph view. It has no package.json and uses no @nodetool-ai packages. See Workflow Runner for the files, protocol and how to run it.