Analyze. Plan. Execute. Without pasting code into a chat.

PlanBefore lets ChatGPT, Claude, Gemini and other AI assistants read the projects you choose, turn a need into small tasks you approve, and hand each one to Codex, your local coding agent. One task, one commit, your review in between. PlanBefore does not upload or mirror your repository.

Your repository is not uploaded or mirrored in PlanBefore. Only the content needed for each request is shared with the assistant you authorized.

curl -fsSL https://planbefore.kvnlabs.xyz/install.sh | sh

Latest release v0.7.0

1 commit

per task, only the files that task changed

0 bytes

of your repository stored by PlanBefore

Start with questions. Add planning and execution when you want them.

  1. Ask your repo anything

    Five minutes. No planning folder, no agent.

    Expose a project and your assistant can browse the tree, read files, search code and symbols, and inspect Git status, diffs and history. Every analysis tool is read-only. Ask "how does login work here?" and get an answer grounded in the real code, not in what you pasted.

    pbnode init
    pbnode login
    pbnode expose ~/Projects/my-project
    pbnode serve --background
    pbnode status
  2. Turn a need into tasks you approve

    Optional. Enable it per project.

    Authorize one folder inside the project and your assistant writes the plan there as Markdown: a requirement, small self-contained tasks with expected behavior and checks, and the review of each one. It stays in your repo, in your Git, in plain text. You approve the requirement before anything is saved and each task before it moves.

    pbnode expose ~/Projects/my-project --spec-root specs
  3. Let your local agent implement them

    Optional. Enable it per machine.

    Approve a task and PlanBefore hands it to the coding agent installed on your machine. The agent gets only that task, nothing else. When it finishes successfully, the Node creates one commit with only the files that task changed, on the requirement’s branch, and returns the result to your assistant for review. No push, no merge, no rewritten history.

    Execution with Codex is enabled separately, when you want the full planning and implementation flow.

    If codex is on your PATH, the Node detects it with nothing else to configure. Only if it is not, point the Node to it:

    # config.yaml
    executor:
      command: codex

Your assistant asks. Your machine answers.

A small app, PlanBefore Node, runs next to your code and keeps an outbound connection to the Hub; no open ports, no VPN. Your assistant talks to the Hub over OAuth 2.1, and the Hub routes each request to the Node that owns the project. Files are read locally, one request at a time.

  1. Your machine

    PlanBefore Node

  2. PlanBefore Hub

    mcp.kvnlabs.xyz/repoflow

  3. Your AI assistant

    ChatGPT, Claude, Gemini or another compatible assistant

Your repository stays on your machine.

PlanBefore does not upload or keep a copy of your repository. Files are read locally and on demand. When the assistant you authorized needs information, only the content required for that request is sent so it can be answered.

  • You choose what to expose. Only the projects you register are reachable; known sensitive files are excluded by default; you can withdraw access whenever you want.
  • Only what each request needs. Files, search results, Git changes or other project information are shared only when the authorized assistant requests them to carry out the analysis.
  • Read-only by default. Analysis tools cannot modify your code. Planning writes go only to the folder you authorize. Execution runs only on tasks you approved.
  • Git with clear limits. During a task the Node owns the branch and the commit: one commit per task, only its files, your previous work untouched. Never push, merge, reset or amend.

Content you share with an AI assistant is also subject to that provider’s policies.

Works with the tools you already have.

Assistants
ChatGPT, Claude, Gemini and other AI assistants.
Execution
Codex today. Follow the releases page for other agents.
Systems
Native builds for macOS, Linux and Windows.

Install PlanBefore with one command

The installer downloads the right build for your system, verifies its SHA-256 checksum and installs pbnode for your user.

macOS / Linux (Terminal)

curl -fsSL https://planbefore.kvnlabs.xyz/install.sh | sh

Windows (PowerShell)

irm https://planbefore.kvnlabs.xyz/install.ps1 | iex

Prefer to do it by hand? The manual installation takes five steps.

Then expose a project and ask your assistant about it.

pbnode login opens the Node approval page on PlanBefore Hub; you sign in with your PlanBefore Auth passkey. The terminal never asks for your credentials.

Read the installation guide