NodlumeLangChain Deep Agents generator
Describe an agent for LangChain Deep Agents, the harness for long-horizon agents on LangGraph, and download the TypeScript project that builds it — createDeepAgent with its system prompt, typed tools, subagents, skills, memory, backend and human-in-the-loop gates. No account, nothing leaves your browser.
The config is the agent
Unlike a filesystem-first framework, nothing here is discovered by path. src/agent.ts is where every tool, subagent and backend decision meets, so the preview opens on it — the other files exist so each piece can be edited on its own.
Validated, never auto-fixed
A tool named Search Web is reported and left out rather than renamed, and one named read_file is refused rather than allowed to silently replace the built-in. A gated tool with no checkpointer is flagged, because that is a runtime throw, not a style preference.
No secrets on the page
The model prefix decides the provider package and the API key name, which is echoed into .env.example. The tool never asks for the value, so there is none to send anywhere.
How to scaffold a LangChain Deep Agent
Four steps from an empty form to a running createDeepAgent project. Every field maps to one option of createDeepAgent, and the preview on the right shows the file it lands in as you type.
Describe the agent
Name it, pick a provider:model id, and write the system prompt and rules. The prefix of the model id decides which provider package and API key the project needs.
Add tools, subagents, skills and memory
Each tool becomes a tool() with a Zod schema; each subagent a spec with its own prompt; each skill a SKILL.md; memory an AGENTS.md the agent edits with edit_file.
Choose the backend and checkpointer
StateBackend keeps files in the thread, FilesystemBackend on disk, LocalShellBackend adds an execute tool. A checkpointer is required the moment any tool pauses for a human.
Download and run
npm install, fill in .env, then npm start -- "your question". The entry point resumes interrupted runs with Command({ resume }) and prints the final message.
What the download contains
- src/agent.ts
- The createDeepAgent call — model, system prompt, tools, subagents, skills, memory, backend, checkpointer, middleware and interruptOn in one place.
- src/system-prompt.md
- The agent's instructions, kept as markdown so they can be edited without touching TypeScript.
- src/tools/*.ts
- One tool() per file with its Zod schema and a stub handler.
- src/subagents/<name>/index.ts
- A SubAgent spec with its own system prompt, optional model, tools and skills.
- skills/<name>/SKILL.md
- Progressive-disclosure skills; the frontmatter name must equal the directory.
- memories/AGENTS.md
- Long-term memory preloaded into the prompt each run, optionally routed to a StoreBackend.
- src/index.ts and src/seed.ts
- A demo invocation with the thread config, the human-in-the-loop resume loop, and — on a StateBackend — the files seeding that makes skills and memory reachable.
Deep Agents, answered
The longer version is in the guide Building LangChain Deep Agents: the config object is the agent. This is an unofficial tool built against the public Deep Agents documentation — the framework itself, and the last word on anything below, lives at docs.langchain.com.
Design the app the agent works for
An agent usually sits behind a product. In the Nodlume workspace you draw that product on a visual structure canvas and export a real Next.js, Tauri, Expo, Ink or extension project from it. The Eve agent generator is this page's sibling for Vercel's filesystem-first framework — the same form, for a runtime that reads its configuration off the folder instead of a config object.