RAAV is local-first product and project memory for teams using Codex, Claude Code, Cursor, and MCP agents. It keeps a shared Product Handbook, task plan, branch lanes, file claims, founder review queue, verification, and audit trail in your repo — zero-LLM by default.

Is RAAV a coding agent?

No. RAAV is persistent product and project memory. Codex, Claude Code, Cursor, and MCP agents do the reasoning and code; RAAV holds product truth, tasks, lanes, claims, review, verification, and audit.

Does RAAV use its own LLM?

No. RAAV is zero-LLM by default. Your coding agents use frontier models; RAAV stores and coordinates the product context they need.

Non-coding solo founders and small teams coordinating multiple AI coding agents on one product repo.

How much does RAAV cost?

Local tier is free with full CLI and local memory. Solo and Team hosted plans are on the waitlist while founder validation completes.

Product memory for Claude Code, Codex, and Cursor

Your agents forget. Your product shouldn't.

RAAV keeps one product memory in your project — the brief, the plan, the decisions, the proof — that every agent reads before it works and updates when it's done. No chat window, no server, no AI credits spent on memory.

Join the waitlist Explore the workspace

Lives in your project · never uses AI credits · yours to keep

Sample workspace · live demo

This is what your agents see. Watch the founder loop play out — or click into the sidebar to explore what needs you, what is building, and what shipped.

Join waitlist

The RAAV founder console — Today view with one decision waiting

Today — What needs you before agents keep working — TodayWhat needs you before agents keep working

Swimlanes — Every lane on a timeline — shipped, building, blocked — SwimlanesEvery lane on a timeline — shipped, building, blocked

Tickets — Scoped work agents can pick up without a repo crawl — TicketsScoped work agents can pick up without a repo crawl

Product memory — The product truth every agent shares — Product memoryThe product truth every agent shares

Review — Approve before anything becomes product truth — ReviewApprove before anything becomes product truth

Today

raav-ai / main · workspacesample data

Today

What needs you before agents keep working

The project is 68% of the way to V1 launch. One decision needs you — nothing else is blocked.

Needs you

Should agent-owned epics require a human reviewer before release?

Codex asked while planning the launch asset. Answering unblocks that lane.

1DECIDE-001 ·codex · 2h

Building now

codex72%

Public launch asset

claude38%

Handbook language for founders

cursor15%

Team roles and seats

Up next

Polish the memory console demo

Scoped to 4 files with acceptance criteria attached — agents read the ticket, not the repo.

everything above is backed by the ledger · verified 09:04 · 64 audit events

Today· autoplay — click to take over

1.0 Understand

See the picture — a landing date nobody typed in

Scope, progress, and a landing date fall out of the work itself — not a status field someone remembered to update. Nobody wrote a status report.

What we are building

RAAV

Persistent product and project memory for agents building your product.

Who it is for

Non-coding solo founders and small teams coordinating Codex, Claude Code, Cursor, and MCP agents.

The promise

Every agent works from the same product truth.

The problem

Coding agents forget. Every new chat re-explains the product, reinvents the roadmap, and contradicts decisions the founder already made.

A day in the life

Before

Tuesday, 09:12. Codex is rewriting the pricing page you froze last week. Claude just invented a fourth persona. Cursor is asking — again — what your product does. Three terminals of scrollback, fourteen explanations, nothing remembered.

After

Tuesday, 09:12. You open Today: one decision needs you, nothing else is blocked. You decide, the lane resumes, and every agent picks up from the same brief — the one you confirmed.

Use cases

Brief a new agent in one command instead of a chat history

confirmed

Approve or reject what agents propose before it becomes truth

confirmed

Run two agents on one sub-project without file collisions

confirmed

Schedule an overnight run and review it at breakfast

inferred

User requirements

Founder-readable views: no PM jargon anywhere

non-functional

confirmed

Agents claim files before editing; conflicts checked

functional

confirmed

Every handoff records files, verification, and risks

functional

confirmed

Progress derived from the ledger — no manual status

functional

confirmed

Works offline, local-first; memory lives in the repo

non-functional

inferred

Dispatch scheduled runs from tickets

functional

needs you

Specification

V1 scope

Product Handbook + review queues
Tickets, lanes, claims, audit
Hosted founder console

Non-goals

Not a coding agent
No LLM calls of its own
Not a general PM tool

Success metrics

Founder answers "can I ship?" unaided
Zero file collisions across parallel agents
New agent productive in one command

Open founder questions

Should agent-owned epics require a human reviewer before release?

Which runner ships first for dispatch: desktop or cloud?

Needs founder review

Scope and promise include inferred assumptions separated from confirmed facts.

How far along we are

V1 ≈ Aug 2

Scope

Started

Done

May 12todayAug 9

Done Scope Projected

codex

18 tickets

claude

12 tickets

cursor

11 tickets

latest update ← DESI-003 handoff: "Today now opens with one sentence and four cards."

2.0 Prioritize

Know what needs you — one decision at a time

Today surfaces the one thing blocked on your call and lets the unblocked work keep running without you. You are the bottleneck on one thing, not all of them.

The project is 68% of the way to V1 launch. One decision needs you — nothing else is blocked.

Needs you

Should agent-owned epics require a human reviewer before release?

Codex asked while planning the launch asset. Answering unblocks that lane.

1DECIDE-001 ·codex · 2h

Building now

codex72%

Public launch asset

claude38%

Handbook language for founders

cursor15%

Team roles and seats

Up next

Polish the memory console demo

Scoped to 4 files with acceptance criteria attached — agents read the ticket, not the repo.

everything above is backed by the ledger · verified 09:04 · 64 audit events

3.0 Direct

Hand agents a scoped ticket, not the whole project

Each ticket carries its scope, its finish line, and its proof — so agents get exactly what they need and read ≈2k tokens instead of crawling the whole project every time.

Every ticket on one timeline. One is waiting on your decision — the rest are moving or queued.

Ticket

Jun 28todayJul 12

State

Workspace designdesign/ws2-workspace-kit

Public launch assetcodex/launch-asset

4.0 Approve

Nothing becomes product truth until you confirm

Agents propose requirements, memory updates, and decisions. RAAV holds them in review until you approve, defer, or reject — assumptions stay out of durable memory until you confirm them.

One decision at a time — answer it, the next one appears.

decision 1 of 3

DECIDE-001needs you

Should every agent-owned epic require a human release reviewer?

This decides how safe multi-agent work feels for non-coding founders.

proposed by codex

5.0 Verify

Two agents in one lane, zero file overlaps

Claims make parallel work safe — see each part of the work on a timeline, even two agents working side by side without stepping on each other or undoing what the other just shipped.

Five sub-projects. Launch asset is waiting on you — at current pace it lands Jul 9.

Sub-project

Jun 28todayJul 12

Status

One memory, in your project

One place that holds the truth about your product.

What it is, what you decided, what's done. Your agents read it before they work and update it when they finish. It lives with your project, on your machine, and it never spends your AI credits — your agents do the thinking, RAAV just remembers.

01Understand

See what the agents believe — confirmed facts, assumptions, open questions.

02Prioritize

Choose the next useful task from goals, requirements, and risks.

03Direct

Send work to Codex, Claude Code, Cursor, or MCP agents in visible lanes.

04Approve

Accept or reject product changes before they become durable truth.

05Verify

Know what changed, what passed, and what is still risky.

The problem

Tuesday, 09:12. Codex is rewriting the pricing page you froze last week. Claude just invented a fourth persona. Cursor is asking — again — what your product does. You've explained it fourteen times, in fourteen chats, and none of them remember.

The problem isn't your agents' code. It's that every chat starts from zero.

How it works

Three steps from a cold repo to shared memory.

Paste one prompt

Set up RAAV in this project.

Copy one line, paste it to your agent, and Claude Code, Codex, or Cursor installs RAAV from inside your project.

Build product memory

RAAV turns repo context, decisions, tasks, and founder input into a Product Handbook with confirmed facts, inferred assumptions, and review gaps.

Coordinate work

Agents read the same overview, claim tasks and files, work in parallel lanes without colliding, and submit their proof back into the history.

Before / after

What changes when agents share memory.

	Without RAAV	With RAAV
Product memory	"Let me explain the app again..."	`raav product handbook` shows confirmed facts, assumptions, and open founder questions
Agent coordination	Two agents can edit the same files blindly	Tasks, branch/worktree lanes, claims, and conflict checks
Founder visibility	Raw terminal logs and scattered chat history	Founder-readable product brief, next decision, risks, and audit trail
Work handoff	Nobody can reconstruct what changed or why	`raav submit` records files, summary, verification, risks, and next action
AI spend	Another product tries to reason with its own model	Zero-LLM by default; Codex and Claude Code do the reasoning

The hard part was never asking an agent to write code. It was keeping the product straight — what we decided, what was still open, what the last agent had already tried. RAAV is that memory. The thing I wished I had at chat number fourteen.

I am Ville Ylläsjärvi. I built FeedSquad solo, with coding agents, from inside the Arctic Circle.

Co-founded Haltian (100+ people, €19M revenue)·Nokia, BMW Designworks·Building from Levi, Finnish Lapland

Save tokens

Save tokens on every agent run.

Most of an agent's tokens go to re-learning your product before it writes a line. RAAV hands it a scoped ticket instead of the whole repo — and never spends your AI credits keeping the memory itself.

Per task an agent picks up

≈2ktokens to read one scoped ticket

The ticket carries its scope, its finish line, and its proof — so the agent reads the context it needs, not the whole codebase, every time it starts.

Per memory update

€0in AI credits — RAAV is zero-LLM

Recording decisions, tasks, and proof is plain bookkeeping. Your agents do the thinking; RAAV just remembers — so your model bill never pays to store what you already know.

You will never pay per agent, or per token.

Start for the price of a coffee, billed by the week, and cancel any time. When you're ready, move to a monthly plan for more projects, longer audit history, and a shared team workspace. You pay for hosted memory and coordination — never for agents, never for usage. Prices exclude taxes.