The product
An execution environment,
not another chat window.
Unifounder runs inside the AI agent you already use. It gives that agent three things it does not have on its own: a structured memory of your business, an order of work, and procedures it has to follow — so what comes back is evidence you can check rather than an answer you have to trust.
This page explains the machine and the words it uses. Every one of them — workflow, stage, agent, memory, gate — means something specific here.
What it gives the agent
Five capabilities, and none of them is a better prompt.
The model you already pay for is not the missing piece. What it lacks is a memory of your business, an order to work in, and procedures it has to follow. Unifounder supplies those three, and everything else on this page is a consequence of them.
Swipe the cardsThe stack cycles on its own — hover to hold one, or read the same five defined in full below.
The vocabulary
Seven words, and the whole system is made of them.
Workflow
One bet, run to a conclusion
A reusable route to an outcome, with a beginning and an end — validating an ICP, building an inbound engine, launching a blog. It runs an ordered set of stages, taken from a blueprint somebody wrote deliberately rather than assembled on the fly.
You can run several at once. They do not merge, and each keeps its own history.
Stage
One question, with a hypothesis
A phase of the workflow. Each stage states the one question it exists to answer and the hypothesis it is testing, in writing, before any work starts — so the result cannot be reinterpreted afterwards into whatever the data happened to say.
It groups the agents below, and it ends at a gate.
Agent
A goal-directed unit of work
One piece of work that plans, acts, observes and reports: audit this surface, choose this audience, design this experiment. A Research Agent discovers and updates memory and hands back a written report — findings, the evidence behind them, and what is still unknown.
Every claim it makes carries a status. Something proven and something assumed do not get to look the same.
Two senses of the word, and they are not the same. Unifounder runs inside the AI agent you already use — Claude Code, Claude Desktop. That is your client. An Agent here is a unit of work your client carries out. One is the worker you brought; the other is the job it picks up.
Action Agent
Work that changes something real
The kind that acts, and the distinction matters. It builds the page, installs the tracking, publishes the post. It hands back records rather than prose — the rows it actually created.
Anything irreversible or public sits inside an Action Agent, behind your explicit approval. A Recurring Agent is the same thing on a cadence; a Follow-up Agent handles a dependency or a question only you can answer.
Company Memory
What the system knows about you
Structured memory, not notes. Your markets, problems, offers, proofs, channels, pages, content and people — as records joined by relationships, read before anything is done and written to when work produces something real.
It is the reason session two does not start from nothing.
Decision gate
Where you decide, not the AI
Every stage ends at one. The system reports what it found, recommends an outcome, and stops. Nothing advances until you choose.
Your choice is recorded with the date and your reasoning — including when you overrule the recommendation.
The gate
Four ways a stage can end. Only one of them throws the work away.
Validate
The next stage opens.
Pivot
This stage runs again, with a new hypothesis.
Invalidate
The workflow returns to an earlier stage.
Kill
The workflow retires, and what you learned stays on record.
The runner-up and what you ruled out stay on record too. A report that shows only the winner has hidden the decision — and six weeks later you will want to know what you rejected and why.
Company Memory
Your company as records, not as a chat transcript.
Markets
who you sell to
Problems
what hurts, in their words
Offers
what you sell
Proofs
what you can actually back up
Channels
how people reach you
Web pages
the surface you control
Content
what you have published
People
who does what
They connect: a problem is felt by a market and answered by an offer; a proof backs a claim and came from a customer. Because the connections are data rather than prose, the agent can follow them — and you can see which records hold up and which are still assumptions.
Where it runs
Inside the agent you already pay for.
Unifounder connects over MCP to the client you already work in — Claude Code, Claude Desktop, or any compatible one. You do not type into a second window, and there is no second model subscription.
Your agent installs skills: plain-text procedures for research, positioning, offers, copy, images, experiments. They are files you can read before you install them, and they are what stops “do some marketing” becoming whatever the model felt like that day.
What it refuses to do
The constraints are the product.
It will not decide for you.
It scores the options, recommends one, and stops.
It will not fill in blanks.
A missing record is reported missing. Assumed stays labelled assumed.
It will not overwrite your call.
If it thinks a number is wrong it writes a note beside it. It does not change it.
It will not publish or send.
Drafts always. Anything irreversible waits for your explicit approval.
It will not invent evidence.
No statistic, quote or customer that does not exist. A gap is reported as a gap.
It will not hide the runner-up.
What the choice ruled out comes attached to it.
Where we actually are
Pre-launch, and not pretending otherwise.
Everything above describes what the software does — that part is checkable, and this site is built and measured with it.
What we cannot yet show you is a founder who says it changed their outcome. Nobody has used it long enough and reported back. That claim is untested, we are not going to dress it up as proven, and testing it honestly is what the waitlist is for.
Free. One email field. No date promised.