TypeScript meta-framework for LangGraph.js

Build LangGraph agents like Next.js apps.

B4.run adds file-system routes, shared and route-local tools, generated types, deterministic tests, durable threads, and build targets around LangGraph.js. Keep the runtime. Drop the boilerplate.

Get started · Migrate from LangGraph.js · Documentation · Discussions

npm create b4-app@latest my-agent

Read the product-loop transcript.

Quickstart

The first activation is the no-key path:

Requires Node.js 24 or later.

npm create b4-app@latest my-agent
cd my-agent
npm install
npm test

In a credential-sanitized clean-room run on August 31, 2026, npm resolved create-b4-app 0.8.21; the generated starter installed and its default fixture suite passed, with one opt-in test file skipped. The observation is version- and date-specific; @latest can move. See the activation receipt.

Why B4.run

• Author the application shape, not route wiring. Put one supported entry in

• Let generated types follow the route. B4.run generates TypeScript route,

• Make the normal test loop deterministic. Fixture-backed tests replay

• Carry one project from local state to build artifacts. Default

How B4.run fits

| Layer | Role | | --- | --- | | LangChain | Remains available. B4.run's built-in agent() path uses LangChain integrations; raw graph and chain routes own their imports and provider behavior, and B4.run does not claim coverage for every LangChain package or provider. | | LangGraph.js | Remains the graph runtime and is required by B4.run. B4.run adds application conventions around it rather than replacing it. | | B4.run | Supplies file-system routing, generated types, local test and development conventions, persistence primitives, and build targets around LangGraph.js. | | Deployment and observability choices | Model providers and LangSmith remain external. B4.run emits a Node runtime and LangSmith artifacts by default, with target-specific options documented separately; it does not provision infrastructure, host the app, or manage secrets. |

What B4.run writes for you

| You author | B4.run discovers or emits | | --- | --- | | One agent, workflow, graph, or chain entry in a route's index.ts | The route identity and runtime entry | | Shared tools in src/tools/ and optional route-local tools in the route's tools/ directory | The tool set available at that route, subject to tool policy and runtime constraints | | Optional state schemas and typed tool signatures | Regenerated route, route-parameter, state, and tool declarations; state exports require discoverable defaults | | Fixture-backed tests and application configuration | A deterministic replay path for those fixtures; live and recording paths stay explicit | | Application source | A runnable Node server, Dockerfile, and LangSmith graph artifacts from the default build |

Already have a graph? Keep its nodes, edges, and imports, expose it as a raw graph route, and validate invocation and checkpointer behavior for each target. The full migration guide covers that incremental path.

What are you building?

• Research assistant
• Chat and workspace assistant
• Memory-backed agent
• Routes and workflows guide

When B4.run fits

• A good fit: a TypeScript team wants LangGraph.js with file-system routes,

• Stay with raw LangGraph.js: if you do not want B4.run's application

• Bring the ecosystem with you: existing raw LangGraph.js graphs can migrate

B4.run is not a hosted AI platform or an infrastructure provisioner. You operate the emitted application or deploy it through a separate platform.

Build with a coding agent

Give your coding agent the framework sources before it writes a route.

Scaffold a new B4.run app and help me build an agent. B4.run is the TypeScript meta-framework for LangGraph — agents and workflows are file-system routes with route-local tools, generated types, and durable threads. Run npm create b4-app@latest my-agent to scaffold, then read https://b4.run/AGENTS.md and https://b4.run/llms-full.txt for the full framework reference before writing any routes.

Run it live

Credentials are provider-specific: the published research starter's OpenAI live path requires OPENAI_API_KEY, while a local Ollama route requires no provider key.

Published @latest (0.8.21)

The clean-room activation recorded above resolved npm @latest to 0.8.21. That published scaffold is a single package. Set its OpenAI key and start its B4.run server on port 3000:

export OPENAI_API_KEY=sk-...
npm run dev

Drive the backend through the Agent Protocol, or follow the Workbench guide to add a browser client; the 0.8.21 scaffold does not include the Workbench package.

In a clean 0.8.21 scaffold inspected on September 1, 2026, npm run build compiles the TypeScript project. That scaffold does not define npm start; use npm run dev for its working server path.

Current source (unreleased 0.8.22)

The current checked-in research template is a two-package workspace with a server and the B4.run Workbench. These commands apply to a scaffold generated from current repository source, not the published @latest package:

npm install
export OPENAI_API_KEY=sk-...
npm run dev:server

In a second terminal:

npm run dev:web

That source template serves the B4.run server on port 3002 and the Workbench on port 3010. Its root workspace defines the production build and start scripts:

npm run build
npm start

Choose and validate the deployment boundary that matches your application: Node, LangSmith, edge targets, or Kubernetes. B4.run emits the artifacts; it does not provision infrastructure, host the application, or manage its secrets.

Maturity and support

B4.run is pre-1.0 and its API surface is moving. Pin versions and read the release notes and upgrade guide. Supported means documented public surfaces on the current release line, not a 1.0 stability or long-term support guarantee.

• Follow SUPPORT.md for support routes, and ask usage questions

• Report security issues through the process in SECURITY.md.
• See CONTRIBUTING.md and CONTRIBUTORS.md before contributing.
• Follow the Code of Conduct.

npm create b4-app@latest my-agent

License

MIT. See LICENSE.