A brief you correct, a customer you agree to
Canopus is an autonomous growth manager that reads your public site and your product events, writes an editable brief of what your product is, and proposes one to three customer segments you validate before any other agent works from them.
You answer twelve questions while the crawl reads your public pages, and the brief and the segments are written from both, with read-only keys for the numbers. Then everything waits: no audit, no channel recommendation and no campaign is written until you have corrected the brief and validated a segment.
How does Canopus learn what my product is?
It reads your landing page and what it links to
One pass, breadth one: the landing page and the pages it links to directly, twelve at most, 400 ms apart, identifying itself as CanopusBot and honouring your robots.txt. It keeps facts rather than HTML: the title, the heading outline, the visible text, the calls to action, the link graph, the JSON-LD and the load time.
An agent writes the brief, with its gaps listed
What the product does, who the site says it is for, the promise, the pricing tiers, the claimed differentiation, the features and the proof. Every feature carries where it was seen, and when PostHog is connected the features are also inferred from your most frequent product events. What could not be worked out goes in an explicit gaps list instead of being invented in a confident-sounding sentence.
You correct it, then validate a customer profile
Five fields are yours to rewrite: the one-liner, what it does, the audience, the promise and the tone. The ICP agent then reads your corrected brief plus aggregate numbers, and proposes one to three segments with personas, pains, objections and value propositions. A version is a proposal until you press validate.
Ideal customerVersion 2Proposed from your product and your existing customers- Solo founder, first paying usersShipped the product, now has to find the people who want it.PainObjectionValue
- Two-person team, flat revenueRevenue stopped moving and nobody owns distribution.ValidatedPainObjectionValue
- Agency owner, productisingTurning a service into a subscription and starting from zero.PainObjectionValue
What does this replace?
Rewriting who you are for, from scratch, in every brief you send a freelance.
One brief, stored, corrected once, read by every agent that writes anything.
Replaces: The positioning doc nobody updatesA persona workshop that produces three slides and no evidence.
Each segment carries the number or the sentence it came from, and a fit rating.
Replaces: A persona templateGuessing which features people actually use from what you remember shipping.
Features are inferred from your most frequent PostHog events, when PostHog is connected.
Replaces: A guessRealising, three campaigns later, that they were written against an old idea of the customer.
Every ICP version is kept, and campaigns record which one they were written from.
Replaces: A shared doc with no history
What runs on its own, and what waits for you?
The crawl runs on its own. It is one pass over a marketing site at breadth one: your landing page, then the pages it links to directly, twelve at most, visited in order of how likely they are to be the pricing page rather than a blog post from 2021. It pauses 400 ms between pages, honours robots.txt, identifies itself as CanopusBot with a URL explaining what it is, and caps each page at 2 MB read from the stream rather than trusting a content-length header. What is stored is facts, never the HTML: the title, the meta description, the heading outline, the visible text, the calls to action, the internal links, alt-text coverage, structured data and the time to last byte. It reads structure and words rather than a rendered page, so it takes no screenshots and says nothing about your design. Twelve pages is right for a marketing site and wrong for a large documentation site, which is a real limit and not a phase.
The crawler refuses to become a way into a private network. This is the one place where you type an address and our server goes and fetches it, which is server-side request forgery in its textbook form, so it has three defences instead of one: the scheme is checked, the hostname is resolved and every returned address compared against the private ranges before a connection is opened, and redirects are followed by hand with every hop re-checked. A public URL that redirects to a cloud metadata endpoint passes the first two checks and is stopped by the third. Canopus also reads your public pages only: it never logs into your application, and a read-only Stripe or PostHog key is the most it ever holds.
The brief is editable because it is a reading, not a record. An agent that read twelve pages knows what your site says; it does not know what you meant, what you are about to ship, or which of two products the page is really about. Five fields are yours: the one-liner, what it does, the audience, the promise and the tone of voice. Your edits are the version every later agent reads, and a later re-crawl refreshes the pages and leaves your text alone. What stays the agent’s is the evidence trail: the features, the pricing tiers and the proof it found, each with where it was seen. When PostHog is connected, features are also inferred from your most frequent product events, with PostHog’s own instrumentation filtered out. That is the closest thing to a real feature list that exists, because it comes from what people do rather than from what a page claims.
The ICP proposes one to three segments, each with a persona, the pains in their words, the objections they will raise, the value propositions against those pains, the triggers that start a search, and where they can be reached. Every segment must carry evidence pointing at a number it was given or a sentence from your site, and a fit rating of strong, possible or speculative. With nothing connected, speculative is the ceiling and the agent is told to say so, because a model asked for an ideal customer profile will produce three beautiful segments for any product on earth. The model never sees an individual customer: it is given counts, medians and distributions, and conversion rates are suppressed below a cohort of ten, because a rate computed on one person is the kind of line that costs a founder a month.
Then everything waits for you. Every run writes a new version, and exactly one version per organization carries the validated flag; validating version four clears it from version three in the same transaction. That flagged version is the only one any downstream agent reads, so an audit, a channel recommendation or a campaign written before you agreed simply does not happen. Editing and validating are separate actions, because fixing a typo is not agreeing to a strategy. Old versions stay, so a campaign can always be read against the profile it was written for. Every model call in the product goes through one function that records the agent, the prompt version, the tokens, the cost and the duration, successful or not, and refuses to start when the organization has reached its monthly budget. No agent has any external side effect: they read and they write rows. Only the campaign publisher sends anything, and only after you approve.
Questions founders ask
See your own numbers first
Connect Stripe and your real revenue is on screen in five minutes, before an agent has said a word. Create your account, answer the questions, and your strategy gets written.