Documentation · Getting started

Getting started, end to end

This is the whole path from a blank idea to code landing in your own repository — greenfield, with links out to each topic's full guide. If you've got 20 minutes, follow it in order.

1Create your project and approve the plan

Click New project, choose Greenfield, and write a paragraph about what you want to build. The platform turns it into an Understanding Document, an engineering Constitution, and a Solution Architecture Document (tech stack, Testing Bar, design system, milestones) — you review and approve each one, in plain language, before the next is generated. Approving the last one decomposes your plan into a real board of tickets.

Full walkthrough: Create your first project.

2Connect a repository

Open Repositories → Connect repository in your project. For a greenfield build, create a fresh empty repository on GitHub first — CogniForge makes the first commit itself, you don't need to seed it with anything. Paste the repo URL and a personal access token (Contents + Pull requests, read/write).

Who clones what: the platform clones the repository itself, server-side, onto whichever machine is doing the building — a build VM you register (step 4), or a background workspace if you're using the VS Code push-channel (step 5). You never run git clone yourself for this, unless you're pointing your own already-open IDE at an existing local checkout of the same repo.

Full walkthrough: Connect a repository.

3Keep the index fresh — how it stays in sync

After every merge the knowledge index (every file, symbol, import relationship, plus semantic embeddings) updates automatically:

Tickets CogniForge implemented — the instant the pull request merges (the platform is watching that PR).
Anything pushed straight to the default branch outside CogniForge's own PRs — a background check notices the remote moved and re-syncs automatically, typically within a couple of minutes.
Want it immediately? Click Sync next to the repository — no need to wait.

This is a deliberate design choice, not a missing feature: CogniForge always dials out to check your repository, and never asks your git provider to call back in with a webhook. That's what lets it work from anywhere — behind a home router, on a laptop with no public IP, on an air-gapped network — with nothing to open or expose.

4Choose how tickets get built

There are three ways to actually implement tickets — pick whichever fits how you work; you can mix them across tickets:

Whichever way you build, you still need a model to power it — see Platform, tenant, or your own subscription for who's responsible for what across the three ways to fund/connect that model.

A build VM — any machine with Docker (even one behind your home router). CogniForge dials out to it, implements tickets there fully unattended, and opens pull requests. See Set up a build VM and Auto-deliver mode.
Your own IDE's agent, driven manually — connect Claude Code, Cursor, Copilot, or Antigravity straight to CogniForge's hosted MCP server and work the queue yourself, ticket by ticket. See Use your IDE's AI.
The CogniForge VS Code extension — the platform pushes a ticket directly into your already-open VS Code workspace and starts a headless Claude Code session there for you. Covered next.

5Connect the VS Code extension (optional)

Generate your own credential — go to your account's API keys page (click your avatar → API keys) and create one. Unlike the MCP integration tokens above, this is self-service: any user can mint their own, no tenant admin required.
Download the extension from Admin → IDE integrations and install the .vsix in VS Code (Extensions panel → … menu → Install from VSIX…).
Open the CogniForge icon in the activity bar, paste your platform URL and API key, and click Connect.
Once connected, click Connect Claude ↔ Platform — this runs claude mcp add for you, so Claude Code's own MCP client is wired up too, with no manual command to copy.
Make sure the repository is actually checked out in whatever folder you have open — the extension delivers tickets into your current VS Code workspace, it doesn't clone one for you (see the note in step 2).

Also needs the Claude Code CLI installed on whatever machine actually runs VS Code's extension host (the remote host, if you're using Remote-SSH) — having the Claude Code extension installed and logged in isn't the same thing. Install it with curl -fsSL https://claude.ai/install.sh | bash.

6Ship it

From here it's the same loop regardless of which delivery mode you picked: a ticket moves through To Do → In Progress → validated → Done, milestone tickets accumulate on a shared branch, and once a milestone's tickets are all done and the whole system passes its end-to-end tests, you get one pull request with the evidence to review and merge. See The board & tickets and Testing & validation.