KARKAS
Get started
📖

Guide

Map the product. Hand it to your agent. Ship.

What it is

kaRkas is the memory your AI tools were missing. You write down what you're building once. Every Cursor chat, every Claude Code session reads it before answering.

That's the whole pitch. The rest of this guide is how to fill it in fast.

Start a project

You have three ways in.

  • From an idea. Type a sentence. We extract title, JTBD, milestones, and a starting stack.
  • From a client interview. Paste the transcript. You get personas and outcomes structured and editable.
  • From a document. Drop a PRD, a Notion export, or a spec. The backbone comes out filled in.

Pick the project type up front if you can — website, mobile, CLI, library. It changes which stack we suggest.

⏱️

A blank project takes ten seconds. Stubbing from text takes about a minute. Either way you land in the same place.

Fill the backbone

Nine sections. You don't have to fill them in order, and you don't have to fill them all. Two or three is usually enough for an agent to feel briefed.

Jobs to be done

Who hires this product, and to do what.

One sentence per persona. The engine reuses these as the north star for everything downstream.

Milestones

MVP, Beta, Scale.

What ships when. Agents check the active milestone before suggesting work.

Functionality matrix

Features versus user states.

Catches the silly gaps. Anonymous user can't reset a password, that kind of thing.

Domain models

Entities and their invariants.

The shape of the data. Optional but agents love it.

Stack

Polyglot components and how they talk.

Inferred from your project type. Edit it before your agent picks a framework for you.

Brand

How the product sounds and looks.

Optional. Even two filled fields help your agent write copy in your voice.

Design

Mockups versus references.

Designs are what you ship. References are what you like. Mark assets current, medium, or outdated.

Tasks

Vertical slices, not chores.

Agents create tasks themselves once the backbone is filled in.

Build

Generated prompts your agent can act on.

Click into a step, copy the prompt, watch the file land in your repo.

Brand & design

Two tabs, both optional, both powerful when filled.

Brand book

Mission. Tone of voice. A palette with hex codes. Typography. Do and don't. Every field is optional. Your agent reads it before writing UI copy.

Design tab

Drop screenshots and Figma exports. Tag each one as a design (what you're shipping) or a reference (inspiration).

  • References get what I like and what I don't like fields. Tell your agent what to copy and what to skip.
  • Mark assets current, medium, or outdated. Outdated assets vanish from the agent's view.
  • Agents can upload their own mockups too. AI uploads carry a badge so you always know what came from a model.

Hand it to your agent

Open the Build tab. Hit Create access token. Copy the curl.

Run it in your project root. The script writes the MCP config, drops your personal skills into the right folder, and your agent is briefed.

  1. 1Pick your agent — Cursor, Claude Code, Cline, Continue, or any MCP client.
  2. 2Paste the curl into a terminal in your repo.
  3. 3Open a new agent chat and ask anything. The first response will reference your real stack, brand, and next task.
🎓

Skills you've saved appear in the install. New Cursor rule? New Claude skill? Drop it in Settings — the next install bundles it.

Chat

Every thread auto-loads the project briefing. You don't paste context. You just ask.

Ask “what did we decide last week about auth?” and your agent searches sibling threads. Ask “draft the onboarding screen copy” and it pulls your brand voice.

Every tool call respects token scopes. Sensitive things, like reading your global memory, are off until you turn them on.

Your model, your bill

Pick the brain per chat. Anthropic, OpenAI, OpenRouter, Ollama, self-hosted — whatever has the keys you have.

We never relay your calls. Keys live in Settings → AI and go straight to the provider you chose.

FAQ

Do I have to fill in everything?

No. Even one filled section beats none. Most users fill JTBD, stack, and a milestone, then let agents help with the rest.

What if my agent ignores the brand book?

It won't, but tighten the tone-of-voice field and add two or three do/don'ts. Specific beats verbose.

Is my data private?

Yes. We don't train on your projects and we don't pipe your chats through anywhere. Self-hosted is coming for full control.

Can my team work in the same project?

Yes. Workspaces share projects. Each member gets their own MCP token and scope set.

Still stuck? Developer docs cover the MCP details.