Closed beta · invites go out in batches
Before an AI agent writes your code, describe the project so it has nothing to guess.
NodePlan turns an idea into a plan: screens, requirements and open questions. It shows what the plan is missing, and you hand the finished plan to the agent that writes the code.
- Free during the beta. Pricing is announced before it ends
- Every request is read by a human
- We reply within 2 working days
Works with Claude Code, Codex and Cursor.
The problem
Your agent guesses where your plan is silent.
You ask an agent for a website with appointment booking. You describe the form and the calendar. The finished site works.
Your description
You
Agent
Done.
- form
- calendar
Agent (thinking)
The plan and the code drift apart, and you only find out in the finished code.
The finished siteit works
The agent left out
Nobody mentioned them in your description, so the agent skipped them or filled them in its own way.
How it works
Three steps from idea to code.
- Step 1
Describe the idea
A few sentences: what you're building, for whom and why. AI proposes screens, requirements and questions, and you accept or reject each one.
- Step 2
See the plan and the gaps
The whole plan on one board. NodePlan shows what's missing: what's required, what's worth adding and how much of the plan is done.
- Step 3
Hand it to your agent
The plan goes to the AI agent that writes the code: Claude Code, Codex or Cursor. When something is missing, it asks in NodePlan and you answer.
You start with a short brief: what it is, who it is for and which problem it solves.
Example
An example: a clinic website with booking.
This is the plan of a demo project. The agent receives the same plan as files.
Smile Clinic — plan
Plan readiness
77%
Screens
- Home
- Services & prices
- Contact & booking
Missing, says the analysis
- 404 page
- Cookie consent
- SEO requirements
Requirements
- Booking 24/7, no phone callmust have
- SMS reminder the day before the visitnice to have
Open question
- Which booking system do we connect?Instead of guessing, the agent asks and waits for your answer.asked by the agent
The agent receives:CLAUDE.mdAGENTS.mdTASKS.mddocs/
This is the start of the real CLAUDE.md of this plan, generated from the plan automatically, without AI.
## 1. Product brief
- **What:** Website for a dental clinic with online appointment booking
- **Type:** Website
- **For whom:** Local patients, parents of young children
- **Problem:** Appointments are booked by phone only and the front desk can't keep up
- **Value / outcome:** A free slot in 30 seconds, no phone call
- **Goals & success metrics:** Half of all visits booked online within three months
## 2. How to use this context
- Specification lives in `docs/` (PRODUCT.md, REQUIREMENTS.md, OPEN-QUESTIONS.md, SCREENS.md, DESIGN.md). Read the doc for a feature before implementing it.
- Every item has a number (§1.3.2) and a stable id (n_…). Reference them in commits and PRs; numbers may shift after regeneration, ids never do.
- Implement only what is defined here. Items marked *(draft)* may change; ⚠️ needs review; ⛔ blocked — do not implement.
- Open questions in OPEN-QUESTIONS.md must NOT be assumed — ask the human.
- Implementation plan: TASKS.md. If the NodePlan MCP server is connected, use it to fetch fresh tasks and report progress (nodeplan_get_tasks, nodeplan_set_status, nodeplan_ask_question).
…Try it on your own project.
Join the betaTechnical
For developers
How this differs from a hand-written CLAUDE.md.
For developers: MCP, Build it, files, tokensFiles from the plan, deterministically
CLAUDE.md, AGENTS.md and TASKS.md are compiled from the plan without AI. The same plan gives the same files, and you regenerate them after the plan changes.
A completeness check before code
Rules for each project type check the plan before the first line of code: what's required, what's recommended.
The agent asks over MCP instead of guessing
The agent fetches tasks, updates their status, asks questions and proposes changes to the plan. You approve them.
When you don't need NodePlan: for a small one-off change, or when you know exactly what to put in CLAUDE.md and keep it current yourself.
Who it's for
Made for people who build with an agent.
A good fit if
- product owners
- agencies
- freelancers
- developers
When not
- NodePlan doesn't write code and doesn't replace the agent.
- You don't need it for a small one-off change in an existing repository.
- When you know exactly what to put in CLAUDE.md and keep it current yourself.
Do I need to know how to code?
Not to plan. The code is written by the agent:
- Claude Code
- Codex
- Cursor
What you can plan
Most mature
- Websitecorporate, landing, e-commerce, media
- Platform / Web appSaaS, CRM, marketplace, booking
- AI agentresearch, support, sales agent
Early version
- Automationform → CRM → AI → email
- Planning without codemarketing strategy, business process
All 9 project types: on the developers page.
Beta
What works in the beta and what doesn't yet.
Send feedback to hello@nodeplan.pro, we read every message.
Solid today
- The graph and the plan editor
- Analysis and plan readiness
- Agent files: CLAUDE.md, TASKS.md, docs/
- The MCP connection: tasks, questions, proposed changes
Early version
- AI features run on free models: a response sometimes comes slowly or fails.
- No-code plans (marketing strategy, business process) and some project types are an early version.
- The editor works best on a computer; on a phone it isn't comfortable yet.
Security
Invitation only
There's no open sign-up: you create an account from an invitation.
Your code stays with you
NodePlan stores the plan, not your code. The repository is yours.
Export and deletion
Download your plan at any time and delete your account in the settings.
FAQ
Frequently asked questions.
How much does it cost and how long does the beta last?
NodePlan is free during the beta. We'll announce pricing in advance, before the beta ends. We haven't announced an end date for the beta yet. You can always export your plan.
How long does a first plan take?
The first plan from a description takes a few minutes. With the required items closed too, it's usually up to 20 minutes in total. You spend that time before the code, not on fixes after it.
What else do I need besides NodePlan?
To plan, just a browser. To build, an AI coding agent: Claude Code, Codex, Cursor or another one. It usually needs its own plan from its vendor (a subscription or pay-as-you-go API). NodePlan doesn't replace or pay for it.
Do I need to know how to code?
Not to plan. NodePlan doesn't write code: the agent does (Claude Code, Codex or Cursor), once you hand it the plan. Starting the agent takes basic skills with it, and the Build it command is pasted into a terminal.
Where does my data go?
NodePlan stores the plan, not your code. Only when you run an AI feature does the NodePlan server send the part of your plan it needs to a model provider. During the beta these are free models via OpenRouter and Groq. They may log requests, and some providers may train on them; data may be processed outside the EEA. So keep confidential details out of the plan if you use AI. Details are in the privacy notice.
When will I get access?
We let people in in batches and reply to every request within 2 working days. In “What do you want to build?” add a few words about the project and the agent you use: it helps us decide the order. Questions: hello@nodeplan.pro.
Access
Join the beta.
Tell us what you want to build: it helps us set the order and help with your first project.
What happens next
- A person reads every request.
- We reply by email within 2 working days.
- You create an account from the invitation and start with a short brief.
- Your details are used only to reply and invite you.
- Already have an account? Sign in