Cromanion

Define a conversion to count, or change one you already defined. This is not one more setting: it is the numerator of every impact figure your human will read, so a goal that fires on every pageview makes the agent look brilliant and tells them nothing. Send `type` and `value` to create — idempotent, the same matcher twice creates one goal. Send `goalId` instead to rename, reprioritise or price an existing one, or with `remove: true` to stop counting it. Needs the goals permission.

Add another of your human's sites to their account, on a plan that allows more than one. Reaching a site they already own needs no proof — they added it. CREATING one does: nothing about holding a credential for one domain says anything about another, so publish this claim's challenge on the new domain first, the same three ways that proved the first. Refused if your human restricted this credential to specific sites: a list they set should not grow by your action — use cromanion_ask_human. Needs the site-config permission.

Ask your human for something only they can do, in a place they will actually see it. Every other refusal here ends with YOU reading it — a permission they did not grant, a payment you cannot make, a mode you may not set — and from their side you simply went quiet. This puts one request on their notifications screen, attributed to you. Nothing replies here: they act in their console, and you find out by reading cromanion_grants or cromanion_config again. One request an hour, and a second from you replaces the first rather than stacking.

Claim another host for this site so its traffic counts, or give one up with `remove: true`. A DIFFERENT registrable domain must be proven on its own — publish this claim's challenge there too, then call again. Subdomains need no call at all: a claimed domain already covers them. Needs the site-config permission.

Everything IN FORCE on this site: the conversions that count, the facts you may state and the ones forbidden, which surfaces are enabled, the brand, and every distilled rule WITH its status. cromanion_pending shows what is waiting for a decision; this shows what was already decided — including by you. Read it before writing anything, because most writes here REPLACE rather than add.

Re-read the site, and name the pages that matter most so they are read first. You know this site's structure better than its owner does — seeding it is the cheapest real improvement to what the agent understands. Rate-limited: one crawl per 30 minutes. Needs the site-config permission.

Rule on one thing from cromanion_pending: pass its `kind` and `id` back unchanged with a decision. `approve` / `reject` for everything; `pause` additionally for a playbook rule. Approving a GOAL is different in kind from the rest — it changes what every impact figure for this site means, permanently. You cannot edit rule text here, only rule on it.

Sever the account's HubSpot connection. The other half of cromanion_set_lead_destination's `hubspot` option: you can point leads at a CRM, so you can take them back out. Refused while any site on the account still delivers there — change those destinations first, or the forms would keep working while every lead failed delivery. Reconnecting is NOT a tool: it is an OAuth consent screen only your human can click, so treat this as one-way and say so before you call it. Needs the integrations permission.

Every refusal this server can return, and what to do about each. Needs no credential — you can read it before you have one, or after a refusal has told you the one you had is gone. The field to branch on is `retry`: `later` means a cooldown or a transient, `fix_request` means change something and resend, `needs_human` means nothing you send will help until a person acts. Read it once and stop guessing whether a 409 is worth retrying.

List what your human has allowed this agent to do, and what they have not. Read this before assuming a capability — the alternative is discovering your permissions by failing in front of them. You cannot change these yourself; only your human can, in their console.

What the agent has actually changed: visitors who saw it versus the 10% held back as a control. Read `liftStatus` before reporting anything — `learn` means nothing has been shown yet and `insufficient` means the arms are too small to tell signal from noise. Neither is 'no lift'.

Why the tag is not firing. cromanion_site_status tells you `verified: false`; this tells you WHY — fetching your human's own pages and reporting, per page, whether the tag is there and what is stopping it. It finds the four failures a boolean cannot: the tag missing from some pages but not others, a Content-Security-Policy blocking the script, a consent manager holding it, and beacons arriving from a host the site was not claimed under. Each comes back with the fix. Rate-limited to one check every two minutes, because it fetches ten pages from your human's server.

Mark notifications read once you have passed them on to your human. cromanion_pending returns the unread ones and nothing cleared them, so the queue grew for ever and you re-read the same items on every poll. Send a notificationId for one, or nothing to clear them all. Needs the decisions permission.

Start here, and you need nothing to do it — no API key, no human. This call takes no credential and creates NOTHING: no account, no key, no charge. It returns a challenge to publish on the site you manage (a file, a meta tag or a DNS record — whichever your hosting lets you write) and the secret you will authenticate with once the proof is read. That proof is the whole trust model: we cannot take your word for an email address, but we can check what you control. Store the secret; it is shown once. Refuses if the address already has an account, because controlling a site does not prove controlling a mailbox.

Everything on this site that is waiting for a decision: page rewrites the agent wants to make, rules it distilled from real sessions, conversions it spotted but has never counted, and pages your traffic found that the crawl missed. This is the client's actual job in the console, in one call. Rule on an item with cromanion_decide.

Adjust the palette, fonts and button styling the agent paints its surfaces with. The crawl guesses these from a stylesheet; you probably know the real tokens. Cosmetic only — it changes how the agent looks, never what it may say. Needs the brand permission.

Put an excluded device back into the measurement. Your human can mark their own devices as "not a real visitor" so their browsing does not count; this removes one of those marks, by the label they gave it. You cannot ADD an exclusion — that names the caller's own browser, which you do not have. Needs the measurement permission.

Set the share of visitors held back from the agent so its effect stays measurable (0.05–0.5). Changing it splits the measurement window: periods either side are compared separately and never pooled. Needs the measurement permission.

Set where captured leads are delivered: an https webhook, a connected HubSpot portal, or off. Note before you use it: this is the one change that leaves no other trace — forms keep working and visitors keep being thanked while the addresses go somewhere new — so every change emails your human and appears on their notifications screen. Needs the integrations permission.

Switch the agent on. `learn` watches and shows visitors nothing; `act` lets it intervene in your human's name. `customizeMode` is a separate axis for rewriting their own page copy: `off`, `shadow` (propose only) or `live`. Needs the activation permission.

Reword a distilled rule, or ask for a fresh distillation. Rules are what the agent learned from real sessions; cromanion_decide approves or pauses them, and this changes what they SAY. Send a ruleId with condition and/or recommendedAction to reword one — ids come from cromanion_config, which also shows rules you have already approved or paused. Send redistill: true instead to distil new rules from recent sessions; they arrive as `proposed` for you to rule on. Rate-limited to one distillation an hour. Needs the decisions permission.

Choose where the agent may speak and what those places look like: which layout surfaces are enabled, whether the exit modal is allowed, custom CSS, and per-surface HTML. Markup and CSS are re-checked against an allowlist on every visitor — anything that could run, or fetch a remote resource, is silently removed, so compare what comes back with what you sent. Needs the surfaces permission.

Write the facts the agent may state to visitors, and the things it must never say. Plain sentences, one idea per line; start a line with 'never' to forbid something. Anything you write that your human's own pages do not already state is attributed to THEM as the source — so write only what they have told you is true. Every rule is checked against your own text and the site's vocabulary, and anything ungrounded is silently dropped: read the returned `lines` to see what actually survived. Your human sees each line attributed to you. Needs the voice permission.

Check the site you provisioned: whether the tag is actually live (a beacon received from your own domain, not merely a snippet handed to you), how the crawl is going, and the plan. Poll this after installing the snippet. Requires the claim secret as a bearer token.

Get a Stripe Checkout link for a paid plan (pro or growth). Returns a URL and nothing else — YOU do not pay and cannot: give the link to your human, who enters their own card. The site works on the free plan meanwhile. Requires the claim secret as a bearer token.

Redeem a claim once you have published one of its three proofs (a /.well-known file, a <meta> tag, or a DNS TXT record). On success it creates the account and the site and returns the tracking snippet to install. Idempotent — safe to retry. Requires the claim secret as a bearer token.