Skip to content
NodePlan
Join the beta

NodePlan

For developers

How a NodePlan plan reaches your agent: the MCP connection, the Build it command, files, project types, tokens, export and security.

MCP

MCP connection

A remote MCP endpoint, https://nodeplan.pro/api/mcp, and a personal token: nothing to install. Works with Claude Code, Codex and Cursor; the Cursor config also fits Windsurf and Gemini CLI.

Claude Code A terminal command

claude mcp add --transport http -s user nodeplan https://nodeplan.pro/api/mcp --header 'Authorization: Bearer np_…'

Codex ~/.codex/config.toml

# 1) ~/.zshrc / ~/.bashrc (or your secret manager):
export NODEPLAN_TOKEN='np_…'

# 2) ~/.codex/config.toml:
[mcp_servers.nodeplan]
url = "https://nodeplan.pro/api/mcp"
bearer_token_env_var = "NODEPLAN_TOKEN"

Cursor ~/.cursor/mcp.json

{
  "mcpServers": {
    "nodeplan": {
      "url": "https://nodeplan.pro/api/mcp",
      "headers": {
        "Authorization": "Bearer np_…"
      }
    }
  }
}

You create the token in NodePlan: project settings, Connect AI agent. It is shown once. Keep it out of the repository: use your user-level config.

What the agent can do

nodeplan_get_context
the project brief and CLAUDE.md: where the agent starts
nodeplan_get_tasks · nodeplan_set_status
tasks in build order; set in progress, done or blocked with a note
nodeplan_ask_question · nodeplan_get_open_questions
a question for the team instead of a guess, and the answers
nodeplan_propose_changes · nodeplan_get_proposals
a proposed change to the spec; you approve it in NodePlan
nodeplan_get_readiness · nodeplan_get_doc · nodeplan_search
plan readiness, generated files, node search

An existing repository: connect NodePlan to your agent there, and it pulls context and tasks, reports progress and asks questions. NodePlan never touches your files: download the generated CLAUDE.md and keep it next to yours, or merge them by hand.

Command

Build it: one command, a new repository

The Build it button in NodePlan gives you a command: it creates a folder, downloads the current plan through a secret link into BUILD-PROMPT.md and starts Claude Code or Codex. For other agents there's a prompt to paste.

mkdir -p clinic && \
  cd clinic && \
  curl -fsSL 'https://nodeplan.pro/api/build/npb_…' -o BUILD-PROMPT.md && \
  claude 'Read BUILD-PROMPT.md fully and follow it step by step.'
  1. The agent creates the repository and saves the plan files in it.
  2. It uses the stack from your brief; if there's none, it picks a mainstream one and notes why in DECISIONS.md.
  3. It works through TASKS.md from top to bottom: build, tests and a commit after every task.
  4. It's instructed to ask before decisions that are hard to undo.

Files

Files for the agent

Files are generated from the plan deterministically, without AI. The same plan always gives the same files. Other project types add their own: docs/FLOWS.md, docs/INTEGRATIONS.md, agents/*.md, STRATEGY.md and more.

CLAUDE.md · AGENTS.md
compact context to start with: the brief first, then a scope map and working rules
TASKS.md
tasks in build order, with the acceptance criteria from your plan
README.md
the project description for the repository
docs/PRODUCT.md
the product: for whom, why, goals
docs/REQUIREMENTS.md
requirements with priority and acceptance criteria
docs/SCREENS.md
screens and their blocks
docs/OPEN-QUESTIONS.md
open questions the agent must not answer on its own
docs/DESIGN.md
the style, when the plan has a design node

Project types

From a booking site to an AI agent.

Each project type has its own blocks, its own completeness rules and its own set of files for the agent. If you're not sure which fits, describe the idea and AI suggests the type. Websites, web apps and AI agents are the most mature today; the rest are marked as an early version.

Build with an AI agent

Claude Code · Codex · Cursor

Plan without code

A board with analysis and a document at the end — early version

Website

Example: Smile ClinicClaude Code · Codex · Cursormost mature

Pages from ready-made sections, the site's goal and audience, analytics, GDPR and the style — the agent builds a site that leads to a booking.

Goal
50% of visits booked online
Actor
Local patients
Screen
Home
Screen
Services & prices
Screen
Contact & booking
Requirement
Booking 24/7, no phone call
Integration
Google Analytics 4
Design
Light, clinical style
In the palette
Home pageAboutServicesContactBlogLanding pagePrivacy policy404 pageBlank pageWebsite goalAudienceAnalytics (GA4 / GTM)and 29 more blocks in the palette
The agent gets
CLAUDE.mdAGENTS.mdTASKS.mdREADME.mddocs/PRODUCT.mddocs/REQUIREMENTS.mddocs/OPEN-QUESTIONS.mddocs/SCREENS.mddocs/DESIGN.md

Plus Other — anything — describe it · Not sure — AI will infer from the description.

Access

Tokens

  • Access: read only, or read and write (status, comments, questions).
  • Every token always expires: 30 days, 90 days or 1 year.
  • A token can be limited to a single project.
  • Revoked in one click on the account page; we store only hashes.
  • Team notes (private) never reach an export, a token or MCP.

Export

Export, no lock-in

You can download the plan at any time: a single file, a ZIP, Markdown with Obsidian links or the full JSON. You delete the account in the settings.

Security

Security

  • Invitation only

    There's no open sign-up: you create an account from an invitation.

  • Protected passwords

    argon2id, a breached-password check and account lockout after repeated failed sign-ins.

  • Scoped agent tokens

    Scope, expiry, one project, revocation: see above.

  • AI only when you ask

    Part of your plan goes to a model provider (OpenRouter, Groq) only when you run an AI feature, and only via the NodePlan server. Data may be processed outside the EEA.

  • Your code stays with you

    NodePlan stores the plan, not your code. Over MCP the agent only sends task statuses, questions, comments and proposed changes.

  • Export and deletion

    Download your plan at any time and delete your account in the settings.

  • Stats only with consent

    Google Analytics on the site turns on only after you agree. No ads, no profiling, and no stats inside the app.

Privacy notice · hello@nodeplan.pro

Comparison

NodePlan next to a prompt and a CLAUDE.md

NodePlan doesn't replace your agent or its plan mode. It comes in earlier, before the first line of code.

  • A long prompt

    When it's enough: A small, one-off change

    What NodePlan adds: A plan you can check, version and hand over again without rewriting

  • A hand-written CLAUDE.md

    When it's enough: You know exactly what to write and keep it current yourself

    What NodePlan adds: Shows what's missing and always generates the files from the current plan

  • Your agent's plan mode

    When it's enough: One task in an existing repo

    What NodePlan adds: A plan for the whole product before the first commit, visible to your team and client, with the agent's questions in one place

  • Spec-driven tools in the terminal

    When it's enough: You work alone, in text files

    What NodePlan adds: A graph instead of files, readable by people who don't code, and completeness rules for 9 project types

Try it on your own project.

Join the beta