fnlowl

Ask for a draft sequence to start sending. Called over the API or MCP this does NOT activate it: it creates a pending approval and returns its requestId, and the sequence starts — enrolling future matching leads and mailing them on its timer — only when a signed-in person approves it in the fnlowl dashboard. Tell the user it is waiting for their approval there. Editing the sequence withdraws a pending request, so call this again after any change. An already-active sequence returns status active and no request.

Ask the fnlowl Advisor a question in plain English. It knows how fnlowl works — widgets and where each kind appears, the embed, sequences, campaigns, sending, credits, plans, MCP — and can read the account's own widgets, leads, analytics, campaigns, sequences and sending status to answer. Suited to "how does X work", "why is Y empty" or "what should I do next". It is read-only and spends an AI credit; a direct read tool is cheaper when the data needed is already known. Takes one self-contained question.

Creates a project — one more site or brand under the account — and returns its id. Pass that id as the project argument on later calls to work inside it. Projects cannot be deleted from here.

Create a transactional template. category is one of authentication, account, orders, notifications. Use {{name}} placeholders in the subject and body for values supplied when it is sent; a placeholder nobody supplies is replaced with nothing. It creates the template only — sending needs a person.

Create a widget in the current project. Six kinds: lead-magnet, popup, survey, announcement-bar, quiz, calculator. It is always created as a draft, and the result lists `issues` — what is still missing before it could go live (empty when nothing is); fix them with update_widget. Going live is a separate step, request_widget_publish. What each kind needs (send only name and kind and the rest starts from sensible defaults): lead-magnet — headline, description, buttonText, contentUrl (a link to the file), optional questions; type "ai" personalises the delivery and needs at least one question. popup — headline, description, buttonText, plus trigger (immediate | exit-intent | scroll | time) with triggerValue (seconds for time, percent for scroll) and placement (center | bottom-right | bottom-left | top-right | top-left); a popup defaults to a time trigger of 8 seconds. For more than one condition, or a click/hover trigger, send triggers (an array of {type: immediate | exit-intent | scroll | time | click | hover, value, selector}; selector is a CSS selector on the host page, only meaningful for click/hover) with triggerMode "any" (the default — opens on the first one met) or "all" (waits for every one). startAt/endAt (epoch milliseconds) bound when a widget runs at all; leave both unset to run with no date limit. survey — questions, each {text, type: short | long | single | multi, options: [] (labels, only for single and multi)}; asks for an email afterwards only if requireEmail is true. quiz — quizQuestions [{text, options: [{label, score}]}] with at least two options each, and quizTiers [{minScore, maxScore, label, message}]; the visitor's total is the sum of the chosen options' scores, and the tiers must cover every possible total from the lowest to the highest with no gap. calculator — calcInputs [{label, min, max, step, default}] (integers), calcFormula using the inputs BY POSITION as a, b, c… with + - * / and brackets (e.g. "a * b * 12"), and calcOutputTemplate where {{result}} is replaced by the answer (e.g. "You could save ${{result}} a year"). announcement-bar — barMessage, barCtaText (the link label), barCtaUrl, barPosition (top | bottom). Popups and bars load sitewide from the one install snippet and can be kept off pages: hideOnValue lists URL pieces they never show on (e.g. "/checkout, /thank-you"), and showOn "url" with showOnValue limits them to pages whose URL contains one of the pieces; showOn "all" is every page. Style is optional and inherits the account's brand when omitted: theme, accentColor, backgroundColor, textColor (hex), cornerStyle, logoUrl.

Create a campaign AS A DRAFT. There is no tool that sends one: sending puts mail in thousands of real inboxes, spends deliverability reputation that cannot be bought back, and cannot be undone. A person sends it from the dashboard. recipientScope is "all" or a single widget id, and unsubscribed or bounced addresses are excluded automatically.

Create a sequence AS A DRAFT. Steps are either an email or a delay in days — a delay is its own step, so two emails can run back to back. A branch may be added at the end; its paths are evaluated first-match-wins, left to right, and a filter with field null always matches, so it only makes sense last. Filters can only test what a lead already carries: unsubscribed and quizTierLabel.

Campaigns in this project, newest first, with their status. Open and click numbers come from the sent-email ledger, not from a counter on the campaign.

List recent email outcomes for this account. Recipients are masked; use this to diagnose queued, delivered, opened, clicked, bounced, complained, or failed mail. `sentToday` counts the last 24 hours on a rolling basis, while `sharedTrial.sendsUsed` counts the account's lifetime sends on the shared trial, so the two legitimately differ. Does not send mail.

List native integration providers and connection health. Credentials are never returned.

Captured leads for the current project, newest first. PAGED — at most 200 per call, and the limit is a hard cap, not a default you can raise; pass offset for the next page (limit 50 with offset 50 returns leads 51 onward). These rows are real people's email addresses. Each lead carries a quality verdict (good, warn, bad, unchecked) from cheap capture-time checks; bad leads are stored but were never mailed, and quality filters the list.

Lists the projects this credential can act on. A project is one site or brand under the account, and every widget, lead, campaign and sequence belongs to one. Every other tool takes an optional project argument: pass one of these ids to run that call in that project. Without it, a call uses the connection's project (the row with current true) or, when no row is current, reads across the whole account. When there is more than one row and none is current, ask the user which website they mean and pass its id on each call.

Who is running through which sequence: for a lead (leadId) or for a sequence (sequenceId), each enrolment's status (active, finished or stopped), the step it is on and when it is next due. This is how to answer "why did this lead get, or not get, the follow-up" — a lead with no enrolment was never enrolled, and activating a sequence enrols future leads, not past ones. At most 100 rows, oldest first.

Sequences in this project with their steps and branch. The shape is BOUNDED: a linear chain of steps, then at most one branch point, each path its own linear chain. No nesting, no rejoining, no parallel paths.

Transactional templates in this account. Built-in starters are marked isCustom false and cannot be deleted.

The same numbers per widget, best-converting first. Use this to answer which magnet is working; use widget_results for where one widget's traffic came from.

Every widget in the current project, newest last. Returns summaries only — id, name, kind, type, status, article, views, completions. Use read_widget for one widget's full content.

Presents the user with a choice between 2 to 4 options, each with a label and a one-line description, one optionally marked recommended. It is for decisions where reasonable options would build different things — which widget kind, what score ranges, what copy or trigger — and nothing the user said settles it. It changes nothing and returns the options for the user to answer.

Move a sequence to draft, which stops new enrolments. Stopping mail is never the slower path, so this applies immediately. Starting one again is activate_sequence, which a person approves: an active sequence enrols every future lead and sends on a timer with nobody watching.

Gives an agent what it needs to advise on a website's lead capture: a snapshot of the website's widgets (kind, live or draft, views, completions, conversion), the six widget kinds with where each appears, and the steps to follow — read the site, ask the user questions with offer_choice, then rank 3-5 fixes. It fetches no pages and changes nothing. Suited to "what should I do on my site" or "audit my site". Scoped to the current website; pass `project` to pick another.

The account's AI context: what the blog is about, its pricing tiers, value props, testimonials, design direction and tone presets. Every AI-personalised lead magnet is generated from exactly this plus the visitor's own answers. It belongs to the whole account and is shared by every website on it, so a connection limited to one website can read it but not change it.

Read recent AI lead-magnet generation outcomes by status and widget. Visitor answers, generated content, lead ids, and failure details are not returned.

Totals for the current project: widget count, page views, completions, conversion percentage, lead count and how many have unsubscribed. Conversion is completions divided by views as a whole percentage; with no views yet it reads 0, which means NO DATA rather than a failure. Completions is not leads — an announcement-bar or an ungated calculator completes without capturing an address. Pass from and to (epoch milliseconds, both) for a date range: the response then also carries that window and the same-length window before it (views, leads, conversions, revenue in cents), plus leads by device and by UTM source. Leads marked bad are excluded from those figures.

The brand: site name, accent colour, sender name, logo and the postal address that every commercial email needs in its footer. With a project in scope this is that website's brand — the account defaults with the website's own overrides applied; without one it is the account defaults.

The A/B experiment a widget is in (or last ran as control): per arm, assigned views, leads, conversion rate and 95% interval in basis points (1234 = 12.34%), and lift against the control arm. result "insufficient_data" means at least one arm has too few views for a rate to mean anything; in that case NO rate or lift is given for any arm, so none is available. A running experiment lists every arm, including one with no views yet. The "holdout" arm sees no widget, so it shows exposure only. beatsControlBp is the chance (in basis points) that a variant beats the control. verdict.status is collecting, leaning or winner, with a leader arm and a plain-language note; it informs a decision and never stops or publishes anything. Views are page loads, so repeat visitors count more than once and the odds are directional.

Read recent queued ESP/webhook handoff outcomes and retry state. Lead identifiers, payloads, and credentials are not returned.

One lead with the answers it gave, and its quiz score and tier if it came from a quiz. The score is frozen at the moment of submission — it is what that visitor actually saw, not a recomputation against the quiz as it stands today.

Reads the account's plan and what it has used of each monthly quota: leads captured, emails sent, AI messages, plus how many websites and widgets exist against the plan's limits. Leads are counted over the billing period and everything else over the calendar month (UTC). `plan` is null when no plan is active. Answers "what is my limit" and "how close am I to the cap". Changes nothing.

Read account sending readiness, domain verification, daily usage, and recent reputation state. While the sending domain is not yet verified it also returns `records`: the DNS records to publish. `sentToday` counts the last 24 hours on a rolling basis, while `sharedTrial.sendsUsed` counts the account's lifetime sends on the shared trial, so the two legitimately differ. Does not send mail.

One widget in full, plus `issues` — what is still missing before it could go live. Includes its questions, quiz questions and options, quiz tiers and calculator inputs. Which of those are meaningful depends on the kind: questions for lead-magnet and survey, quiz fields for quiz, calculator inputs for calculator, none for popup and announcement-bar. `placement` says where it shows on a site: popups and bars run on every page, every other kind only where a page embeds it — going live alone does not place it on any page.

One widget's views, completions, conversion, and its leads grouped by referrer. The referrer is whatever the visitor's browser reported when they submitted, so it is a rough attribution, not a tracked source. Also returns the funnel (views, completions, leads, conversions, revenue in cents; conversions are purchases credited to a lead within 14 days of capture), questionDropoff (how many visitors left a survey on each question, in question order) and the same optional from/to date range as read_analytics.

Ask for a draft widget to go live. This requests publication; it does not publish. Called over the API or MCP with a key it does NOT publish: it creates a pending approval carrying a before/after snapshot and returns it, and the widget goes live only when a signed-in person approves it in the dashboard. The result also says where the widget will appear (`placement` — going live does not put a quiz, survey, calculator or lead magnet on any page by itself) and lists `warnings`, such as copy that promises an email no active sequence will send.

Split traffic between a control widget and 1 to 3 variants. The variants must already exist as their own live widgets, of the same kind and project as the control (created with create_widget and made live with request_widget_publish). On a small site one variant at a time is the practical limit: every extra arm needs the same traffic again. Traffic splits evenly unless trafficPcts is given: one whole-number percentage per arm, the control first and then each variant in the order listed, which together with holdoutPct must add up to 100. holdoutPct (0 to 50, default 0) is an optional share that sees no widget; it can never produce a lead, so it shows exposure only and does NOT measure lift. read_experiment_results reports views, leads, rate and lift per arm, plus a verdict, once it is running.

End a running experiment. Counts stop with it; starting a new one on the same widgets begins with fresh counts, not a resume.

Change the AI context: the about text, design direction, value props, pricing tiers and testimonials that every AI-personalised lead magnet is generated from. Changes apply to generations from then on; lead magnets already sent are not rewritten. Only the fields sent are changed. It holds what the site actually says, so testimonials and prices come from the owner. The AI context belongs to the whole account and is shared by every website on it, so it cannot be changed from a connection limited to one website.

Change brand settings. Only the fields sent are changed. With a project in scope this sets that website's own override and leaves the account defaults — and every other website — alone; without one it changes the account defaults every website inherits. The accent colour must be a hex value — it is interpolated into a stylesheet that renders on customers' websites.

Change a draft campaign. A campaign that has already been sent cannot be edited — the ledger describes what actually went out, and editing it afterwards would make that a lie.

Change a DRAFT sequence. Only the fields sent are changed, except that sending steps or branch replaces the whole of that part, so send the complete chain you want. It cannot edit an active sequence and cannot activate one: an active sequence is paused with pause_sequence, edited, and started again with activate_sequence. Enrolments already running keep the plan they started with.

Change a template's name, category, subject or body — only the fields sent. Built-in starters can be rewritten too (they just cannot be deleted). It changes what later sends of this template say; it sends nothing.

Change a widget. Only the fields sent are changed, with one exception: sending ANY of questions, quizQuestions, quizTiers or calcInputs replaces ALL FOUR with exactly what is sent, so send the complete set you want to end up with. Publishing is a separate action. Views and completions cannot be set — they are counted from real traffic. The result lists `issues` — what is still missing before it could go live. What each kind needs (send only name and kind and the rest starts from sensible defaults): lead-magnet — headline, description, buttonText, contentUrl (a link to the file), optional questions; type "ai" personalises the delivery and needs at least one question. popup — headline, description, buttonText, plus trigger (immediate | exit-intent | scroll | time) with triggerValue (seconds for time, percent for scroll) and placement (center | bottom-right | bottom-left | top-right | top-left); a popup defaults to a time trigger of 8 seconds. For more than one condition, or a click/hover trigger, send triggers (an array of {type: immediate | exit-intent | scroll | time | click | hover, value, selector}; selector is a CSS selector on the host page, only meaningful for click/hover) with triggerMode "any" (the default — opens on the first one met) or "all" (waits for every one). startAt/endAt (epoch milliseconds) bound when a widget runs at all; leave both unset to run with no date limit. survey — questions, each {text, type: short | long | single | multi, options: [] (labels, only for single and multi)}; asks for an email afterwards only if requireEmail is true. quiz — quizQuestions [{text, options: [{label, score}]}] with at least two options each, and quizTiers [{minScore, maxScore, label, message}]; the visitor's total is the sum of the chosen options' scores, and the tiers must cover every possible total from the lowest to the highest with no gap. calculator — calcInputs [{label, min, max, step, default}] (integers), calcFormula using the inputs BY POSITION as a, b, c… with + - * / and brackets (e.g. "a * b * 12"), and calcOutputTemplate where {{result}} is replaced by the answer (e.g. "You could save ${{result}} a year"). announcement-bar — barMessage, barCtaText (the link label), barCtaUrl, barPosition (top | bottom). Popups and bars load sitewide from the one install snippet and can be kept off pages: hideOnValue lists URL pieces they never show on (e.g. "/checkout, /thank-you"), and showOn "url" with showOnValue limits them to pages whose URL contains one of the pieces; showOn "all" is every page. Style is optional and inherits the account's brand when omitted: theme, accentColor, backgroundColor, textColor (hex), cornerStyle, logoUrl.