Persona Console

Long records and smaller screens

Long record references wrap inside the role workspace. Wider tables scroll within their panel so navigation and surrounding controls remain on screen.

Page layout

The shared header keeps Today, Activity, Work, Team, and Settings in one row. Search opens the command palette. The account button contains display settings, help, and Explore this workspace for additional destinations.

Workspaces and saved work

Specialist pages show the entitled role's latest saved research and drafts beside its brief. Research opens saved research. Drafts opens Work with this role selected, including its research and drafts. History opens its Activity records. Role settings opens the console. An incomplete source warning means results are missing; it does not mean the agent has never produced work. Only work belonging to your account appears in this summary.

When your mailbox is paused

If too much mail from your Gmail or Outlook bounced this week, approving a draft to send by email from that mailbox, and the Gmail draft button, are disabled, and hovering them gives the reason. Check the recipient, then choose I checked this address beside the button to clear that one draft. Every other draft stays held until you resume the mailbox in Connections.

Open the original result

Activity and Work can open a specific draft or research source directly. The Drafts tab opens the requested saved record even when it is outside the recent queue. Historical drafts show their status and disable editing, snoozing and dismissal when those actions are unavailable. A draft held for later (a win-back timed for two weeks after a cancellation, a payment email waiting for the card retry, or one you snoozed) can be opened and read by its link, but it shows "Held until" with the date, and its Gmail and Salesforce buttons stay disabled until then. Approving it anywhere before that date is refused and nothing is sent. Snoozing a held draft can only move its date later, and dismissing then restoring it keeps the date. The Library opens the selected source with its facts, quotes and sales application. Existing approval and export controls remain in the original workspace.

Overview

Every role (gtmcron, salescron, leadercron, secronjob, partnercron, bdrcron, crocronjob, cscron, supportcron) has a console at /<persona>/console. It is the same control-plane shell, branded to each persona. The console never runs agents itself. It drives the hosted runtime that runs headless on our servers on a schedule, so nothing depends on you keeping an app open. The cross-product landing is /home.

Scoped to you. Your key, budget, schedule, and drafts are private to your account. The work runs on our servers, so we process the data you connect: it is encrypted, never sold, never used to train a model, and you can revoke access anytime.

The frame (2026-09-11)

The console opens with the role mark, the dot in the role's accent and its name that Today, Activity and the Work panels use, then the tab strip as a row with a hairline under it. The tabs, tiles, tables and buttons inside take the console's own styles in both themes; before this pass the rules written for them were scoped to the page template and never reached what the console draws at runtime, so a button rendered as the browser's default. The accent text step now follows the theme rather than painting the light step on the dark ground. Tables read from the left edge inside the centered column.

Inside the tabs (2026-09-12). What each tab draws now follows the same three rules as the rest of the app. A read-only summary, your goals and pace, the cohort medians, the team rollup, is a head, a hairline and rows, so the panel no longer stacks a card inside the card it already sits in. A card is kept for the blocks you act in: the roleplay and brief pickers, the proof-point form. The role mark replaces the grey eyebrow on the things the role produced, your quarter recap, the week's one skill, a roleplay scorecard, the open findings, so you can see at a glance which of them came from the agent and which are your own settings. Every text box and input takes the theme instead of the browser default, which on the dark ground was a grey slab with a placeholder you could not read.

This hire's number

Under the heading, the console shows the one number this role was hired to move, as a sentence: its baseline from the first week of connected data after the hire, the readings at day 30 and day 60, and your target. When a week had no data it says so and names the connection that supplies it. The line is read-only here; pick the number, change the target and choose whether to show the result on /watch from Home, which explains each role's list and where its data comes from.

How it runs

The hosted runtime runs on our servers on a schedule. There is nothing to install and no app to keep open. Deterministic detections are free and always available. AI drafting runs on your own provider key: every tier is bring-your-own-key, and the AI tier turns on once you add a key in Settings. Without one you still get the full deterministic tier at no cost. Where the deployment turns it on, a small platform-funded allowance covers your first runs before a key is needed. A key turns drafting on; nothing is sent until you turn the send layer on.

When the included budget runs out. AI runs pause until the reset at the start of next month, and every deterministic check carries on unchanged, so the ranking and the findings keep arriving. The Settings tab says so where the meter is and offers the one thing that changes it: add your own provider key and runs go on your key, metered to you and not capped by us. The header pill reads "AI paused" so a quiet month is never a mystery. It used to say only a zone name, and the blocked-run message pointed at raising a cap, which is an operator setting no control on that page can move.

Some work is about one account. A morning brief reads the whole book and produces one draft. A pre-call brief is about a particular company, and those skills now run once per account: each run names its account, carries its own key, and produces its own draft, so two accounts on the same day are two briefs rather than one with no company on it. The agent is told which account it is working on, with the name, domain and last-worked date already on file, and is asked to work only from that account's records. A run you start for one account from the console gets the same block. Before 12 October 2026 a scheduled run named the account only on its draft, so the agent wrote every account's brief from the same prompt. The rotation takes the least recently worked accounts first, up to three per run, and an account that already had that skill today is skipped so a retry or a second dispatch cannot produce a duplicate. An account your exclusion policy holds is skipped with the rule named, and a role with no accounts to work on records a skip that says so rather than writing a brief about nobody. On the Team page you can schedule a role on up to three chosen accounts; from the time you pick, those accounts go first in the rotation and still pass every other check.

Draft-first. A run only ever creates a draft you approve. Nothing is sent without your yes, and you choose where each draft lands in Settings: the in-app Draft queue, which is always on, or your Google Drive. Gmail drafts and Outlook drafts are chosen for every role at once on Home, and only while that mailbox is connected. Once mailbox delivery is switched on (it stays off until the app passes a Google security review, not yet started), a role whose destination is Gmail drafts also places each scheduled draft that is a message (not a brief or a report) in your Gmail Drafts folder, unsent and with no recipient. It stays listed here, but you edit and send it from Gmail: editing it here or approving it to send by email is refused with the reason, and → Gmail draft on it returns the same draft rather than a second copy. A message leaves only when you choose to send it, and only once the send layer is on.

Where this role's work lands is your answer, not a guess. Settings carries three choices: read it in the app, have it written to a Google Drive folder, or send it to your own self-hosted kernel. The queue is always on, so before September 2026 there was no way to tell a founder who had chosen it from one who had not answered, and the setup step asking where output goes stayed open for the life of the account. It is recorded per role, so one role can write to Drive while another stays in the app. Changing it never moves work already queued: what was queued keeps going where it was going, and the role card says so while both are in flight. An account that had Drive delivery or a kernel before this shipped keeps it and is asked to confirm it once, because a flag somebody turned on is not the same as an answer they gave. A destination that stops being writable turns back into a question rather than leaving the step green over a dead target, and that includes a Drive folder whose last delivery did not arrive: the toggle staying on after a revoked Google token is exactly the case where the product would otherwise keep telling you your work was being filed.

The answer is the last word on where work goes, not a preference shown back to you. Choosing the app stops Drive delivery even if the Drive toggle is still on, so the console and your Drive folder cannot disagree about where your account material is going.

Setup and readiness. The console's checklist counts your own records only; the sample book never marks "Connect a data source" done, and the same rule runs on Home so the two agree. "Choose where output goes" is done when you have picked one of the three and it is writable. Home also shows readiness per role, derived from what each role's free checks read, so this console's role can be waiting on a record type while another role is live.

Which tab opens. The console opens on Runs as soon as the role has a run to show, so the first thing you see is what it did. While nothing has run it opens on Schedule, where the setup checklist and the free checks are. A link that names a tab (?tab=runs from Home, for example) still opens that tab, and the sample-data link still opens Settings.

Before anything has run, the Drafts and Runs tabs do not just say "nothing here": they point you at the next step, turn the runtime on from the Schedule tab, run a skill now, run the free checks, or load sample data in Settings to try the review flow before you connect live data.

The tabs

Seven tabs are always there: Schedule, Draft queue, Library, Runs, Handoffs, Build, and Settings. Which of the rest you get follows your seat. A leadership seat (gtmcron, leadercron, partnercron, crocronjob) adds Team, for eight. A front-line seat (salescron, secronjob, bdrcron, cscron, supportcron) adds Coaching and Live coach, for nine.

• Schedule. A master runtime on/off switch, plus every skill with its cadence and tier. Toggle a skill on or off, or run one right now. Deterministic skills are free; AI skills run under your budget cap. Until you are set up, a compact "finish setup: N of 5 ready" card sits at the top with the exact next step; it disappears once all five are done. The five, in the order you should do them: tell your roles about the business, add your AI provider key, connect a data source, choose where output goes, turn the runtime on. Each one links to the place it is actually finished, and for the three that happen here that is the pane holding the control rather than the console's front door, because the console opens whatever tab a link asks for (?tab=settings, ?tab=schedule). Run the free checks now runs the AI-free deal-risk checks on demand and shows the findings inline, so you get value the moment you connect data (or load sample data) without waiting for the next scheduled pass. Every product also includes a lightweight score-narration skill that reads your role's deterministic score board (deal risk, health, awareness, ecosystem fit, technical risk, attention risk, or efficiency, whichever fits the seat) and drafts the plain-language "why" behind the ranking, the number stays deterministic, the AI tier only narrates it. The runtime checks in three times a day (morning, midday, evening), so a skill can run at the right time, for example a morning brief, a midday deal-risk sweep, and an end-of-day follow-up, all from one schedule. Your seat's headline skill always carries a cadence. Most skills are things you ask for, and they stay manual on purpose so your budget goes where you point it; the one the role is named after is the exception, because it is the one that should be waiting for you. Three seats did not have that until September 2026: a BDR's speed-to-lead, an AE's value-framing brief and an SE's discovery synthesis all sat manual, so the headline of each of those products only ran when somebody pressed a button. Speed-to-lead now runs at midday, and the other two weekly, on Wednesday and Monday.
• Draft queue. At the top, five performance tiles computed over your whole outreach log for the window you pick (24h / 7d / 30d / 90d / all): Drafted, Sent, % Sent, Replied, and Reply %. Replies are counted honestly, only a real human reply counts; an out-of-office auto-reply or a bounce never inflates your reply rate. If you run a self-hosted (Lane B) kernel, its LLM spend shows as UNKNOWN here, because it runs on your own key inside your n8n, never $0. Below the tiles, the drafts your hosted runs produced: open one to edit it inline, then Copy it or send it on. Every draft carries a Why this draft line, the skill that wrote it and the exact sources it was allowed to cite, plus a flag if any stat went uncited (and the exact unsupported claims listed out, so you verify those numbers before sending), so you can trust and defend it before it goes out. Your saved edit is what gets delivered: with managed connectors on, → Gmail draft lands your edited version as a draft in your real mailbox and → SF task logs a Salesforce task to review. Nothing is ever auto-sent. For a leadership draft (a scorecard, a QBR, or a board pack), → Drive exports that draft to your Drive "Accounts" folder as a Google Doc, so you can bring it to the meeting. It uses the same Drive connection and opt-in as the automatic delivery, so turn on Drive delivery in Settings and connect Google first. Print / PDF opens any draft as a clean printable page in a new tab, where your browser's print dialog saves it as a PDF, no connection needed. A deck draft carries a second link, Slides, which lays the same draft out one slide per section for showing a buyer rather than one page to read. Both print from the saved draft, so save your edit first. When you Dismiss a draft, a one-tap chooser asks why (not relevant, wrong facts, off tone, or just dismiss); the reason is captured to teach the seat's craft library over time, so the system stops repeating a draft you already rejected. It is optional, and "just dismiss" records nothing. Right after you dismiss, a Restore appears in case it was a mistake. You can also Snooze a draft to hide it for three days; it returns to the queue on its own. A daily learn step reads those reasons: when the same skill keeps getting dismissed for the same reason, it drops a short library review suggestion into your queue (for example, "recent pre-call briefs were dismissed as wrong facts, add a proof point"). It is a suggestion only, points you at your Library tab, and never edits anything for you.
• Library. The two layers that ground every AI draft: your seat's built-in craft library and the proof points you add yourself. Review either, add or remove your own, and drop any built-in craft you do not want. "Your knowledge library (grounding)" below covers what each layer is allowed to do.
• Runs. This role's open findings first, then what ran. The findings are everything this role has raised and you have not answered; the highest few of them are ranked against your other roles on Home, and this is the full list, so a finding that did not make today's line is still reachable. Arriving from a link on Home, the finding you clicked is drawn first and marked. A sample finding says so, the same as on Home. Under them, what ran, when, on which tier, and the outcome, read from the run ledger. A run that fails says why: open it and the reason is on the record, including the case where the model provider stopped responding. Every AI run is bounded, one request stops at 150 seconds and the whole run stops at 240, so a provider that hangs shows up here as a failed run with a timeout reason rather than a window where nothing appears to have happened. The next scheduled pass tries again.
• Build. Two read-only tools for extending what your runtime can do. Connected servers lists the MCP servers your runtime can reach using a credential you have already authorised, and Probe asks one whether it answers and what tools it offers. A probe lists tools and stops: it never calls one, because invoking your tool to prove it could invoke your tool would have done something on your behalf that nobody asked for. If a server needs your own instance URL (Salesforce), it asks for it rather than guessing. Validate a skill draft checks a custom SKILL.md against the same compiler the hosted fleet runs, so a draft that passes here is one your runtime can run. It answers with either every objection to the draft (each with the rule, what is wrong, and why that rule exists) or the manifest entry it would become, its tier, any config tokens it expects that your tenant has not set, and the first part of the compiled prompt. It saves nothing: authoring is feedback here, and saving is a separate decision with its own governance.
• Coaching (front-line seats only). Your own weekly card: where you improved (activity and reliability, week over week), the one skill to work on (your highest open gap this period, phrased as a skill), and the exact play from the craft library to run for it. The play is not just named: open "Show the play" on the card to read its full trigger, steps, and success criterion inline, pulled straight from the craft library. Deterministic and token-free, built from your own open deals and run activity, so it costs no AI budget. If your open deals are all clean, it says so. It also shows your goals and pace: each active goal with whether you are ahead, on pace, or behind for the week or month, and for a goal you are behind on, the tool to run to catch up. Set goals on the Activity page. When enough peers opt in, it also shows a "Reps like you" line: how your draft, meeting, agent-run, and action counts compare to the anonymized cohort median. It is opt-in, shows only aggregate medians, and stays hidden below 5 peers so no one can be identified. Opt out on the Activity page. At the bottom, a "Your quarter" recap tells the story of your quarter in your own numbers: deals won and deals advanced a stage (from the real stage-history timeline), plus agent runs on your behalf, drafts prepared, and actions taken in your tools. It counts only what is tracked, so it is honest, and it costs no AI budget. On salescron, at the start of each calendar quarter the runtime also drafts a warm "Your quarter" narrative (under your budget): the same numbers told as human prose you would be proud to read, grounded in your own accounts and opps. It lands in the draft queue for your review, exportable to Drive or print, never sent.
• Team (leadership seats only). On a leadership seat (gtmcron, leadercron, partnercron, crocronjob), a manager Home that answers "who needs help and why": a rep scorecard ranked by activity, plus coaching flags for anyone silent or erroring this period. The read is metadata only and enforced by the database, you see that a rep ran, when, and the outcome, never what they wrote, and only for teams you own. It costs no AI budget. A leader who owns no team sees an empty state until reps start running. When your reps' pipeline has been summarized, the rollup also shows a team forecast (committed plus weighted pipeline against quota) and a coverage read (how far the weighted pipeline covers the gap to plan). That read stays metadata only: it sums a per-rep, numbers-only pipeline summary (open pipeline, committed, quota, win rate), never a rep's deals or notes. Alongside this deterministic view, leadercron now also runs a weekly AI coaching note (Wednesday, under your budget): it reads how deals are moving in the book and drafts the pattern, the one behavior to change, and a practice rep, so you have it before your 1:1s. It lands in the draft queue for your review, never sent.
• Live coach (front-line seats only). Interactive coaching on demand, alongside the passive weekly Coaching card. Six tools in one tab: Roleplay practice runs a live buyer scenario (pick a scenario, hold a back-and-forth with an AI buyer, get a mid-scene hint and an end-of-scene scorecard); Meeting prep builds a pre-call brief for a meeting from your connected calendar (posture, goals, questions to ask, likely objections, the closing play), grounded in your seat's sales playbooks; Inbound reply drafts a reply to a recent inbox thread; Cadence shows the multi-touch sequences the daily run designed for opps your inbox triage flagged as at risk, with the reasoning for the shape and each step's channel, timing, message and what it is looking for. Accept a step to mark it done so the next run knows where you are, or skip it. Battle card writes a deal-specific card for a competitor you have on file: what to listen for, the concerns that come up at the stage you name, the differentiators that land with the persona you name, one landmine, and one next move. You can add a competitor from the same tab. Loss postmortem takes a deal you closed and drafts the lesson from it, reading the close outcome, the last forecast before it closed and the patterns across your other deals; it is saved against the deal, and a theme that appears in three or more of your postmortems becomes a candidate takeaway for review. Nothing is sent from here either. Meeting prep and Inbound reply end with a Grounded in line naming the playbooks the answer was built from. Click one to read the passage it drew on, so you can check the advice against the craft behind it rather than taking it on trust. If you have rewritten one of those playbooks for your own agent, the line names your version and the tools read it: since 2026-09-06 a rewrite reaches every tool on this tab, the scheduled runs and the inline assistant alike, rather than the inline assistant alone. Each runs server-side on your own key when you have saved one in Settings (BYOK, metered to you), otherwise on the hosted key under your budget. If neither key is set, the tool does not dead-end: it hands you the ready-to-run prompt with a copy button so you can run it in any AI chat. If your saved key is there but can no longer be read, these tools say so and tell you to paste it again, rather than reporting that no key is set and sending you to add one you already have. Always draft-only, nothing is ever sent. The pickers list what you have connected (or your sample data); connect your calendar and inbox on the Connect page to fill them.
• Handoffs. Work another role sends to this role appears with its source signal, source run, entity, decision, due date and expected artifact. A multi-role chain records every stage before the first stage starts. Each stage keeps the same correlation ID and source version, names the person who owns it, and waits for the prior stage to close. Choose Accept handoff to record that this role has taken open work; the handoff then shows when it was accepted and stays open until you complete or decline it. The Team page lists the same open handoffs across your roles with the same Accept control. An open handoff can be completed or declined only after you record the decision and resulting work. Completing a chain stage also attaches the record it produced, and each stage names the record type it needs: a confirmed profile for GTM, a sourced contact and an approved (sent) draft for Sales development, an opportunity for Sales, a cited answer for the SE, a partner introduction and then a partner attributed opportunity for CHAIN3, an outreach hold and later its release for CHAIN4, a customer permission and then a Studio content draft for CHAIN5, a capacity or budget plan for the Revenue leader stage of CHAIN6, an experiment or hiring plan for the GTM stage of CHAIN7, and the founder decision on founder stages. Pick the record from your own records in the stage's select. Complete is refused with a plain reason when a required record is missing, and a record that is not yours reads the same as one that does not exist. Declining still needs only a reason. Five stages make their record in place: the SE stage records a cited answer from the buyer question, the deal and entries from your answer library or product gaps, with no AI call (a product gap is listed by its theme, and the answer keeps the gap's theme and note as its cited excerpt); the Partner stage records the approved introduction to one of your accounts; the CS stage records who at the customer gave permission, when, and what it covers; and the CHAIN6 Revenue leader stage and the CHAIN7 GTM stage record their plan: a title, the plan type, the decision it answers, the owner role, the amount or headcount where given, a review date and an optional measure. A budget plan needs its amount and a hiring plan its number of hires, the review date cannot be in the past, and an account plan cannot close either stage. The founder decision page in Team decisions shows the plan in full: the CHAIN7 proposal before the founder decides, and the CHAIN6 plan once the Revenue leader records it. The Sales stage of CHAIN3 opens or links the opportunity with the partner and introduction written onto the deal, and the GTM stage of CHAIN5 creates a Studio draft only under the permission the CS stage attached. The next stage shows the attached records as evidence from the preceding stage, and a closed stage lists its outputs beside its outcome. Closing it opens its dependent stage once. Declining it marks later waiting stages unresolved. Work with a missed deadline becomes expired; daily maintenance also marks its waiting descendants unresolved. Retrying the same source version preserves recorded decisions. Cross-user assignment is not supported by this user-scoped queue. The initiating role can start a chain from one of its recent real findings with an account or opportunity reference. The source selector shows the account or opportunity reference and recorded time so repeated finding titles can be distinguished. The server resolves its source version and retains the reviewed evidence. Starting records work to do, not a completed first output. All participating roles need entitlements. Ready founder stages appear in Today and open a selected source and preceding-outcome review in Team decisions. The founder can edit the proposed wording, use AI Assist and record or decline the decision. A missing source snapshot or unfinished preceding stage prevents a decision. Live source-to-outcome acceptance remains separate from these controls. Completed, declined, expired and unresolved work keeps its recorded outcome. Daily maintenance removes terminal history older than 18 months in bounded batches, after later waiting stages have been resolved. A failed read says it could not load rather than showing an empty queue.
• Settings. Your budget/usage meter, your identity (fills the {{TOKEN}} values in the agents), bring-your-own-key, managed connectors, where drafts land (the in-app queue, or Google Drive), and sample data. The identity fields are the same ones as Account & Settings › Profile: one store, either door, and a change in one shows up in the other. It is shared across all your consoles rather than set per role, because your name and title do not change between them.

What the tabs are fed by

• Your confirmed business brief, including its narrative. Every skill on every seat reads the confirmed brief: who you are, what you sell, who you sell to. Since October 2026 that includes the narrative: your conviction, what your reader is against, who they become, and up to three stories from your story bank, so a week of drafts argues one belief instead of a new idea each run. A draft brief, or a narrative you have not confirmed, reaches no skill. On gtmcron, the messaging experiment tracker uses the conviction as its experiment: it proposes six to eight wordings of it and reports which ones people quote back. When the brief names a market motion (lighthouse, landgrab or signal-led), every skill also reads what it means for list size, how a first email opens, pace, and the next step to propose. A brief with no motion reads exactly as before.
• Sample data. Not ready to connect live data yet? In Settings, Load sample data seeds a small sample book so the free deterministic checks surface real findings (a slipped close date, an overdue next step, an unactioned account signal) and the draft queue shows the review and "why this draft" experience. Everything is clearly marked sample, and Clear sample data removes exactly it and nothing else.
• Role checks on your own data (free). Beyond the generic checks, each seat's own analysis engines run on the same schedule over your stored records, with no AI spend. For an AE, the free sweep scores each open deal's risk with the MEDDPICC gap driving it, flags deals stuck in a stage with the move to advance, and names the buying-committee white space on accounts you are already working. For a CSM, it surfaces at-risk accounts with the churn drivers, renewals coming due on accounts that are not healthy, accounts whose realized value is stalling, accounts being served below the motion their ARR and health earn, and healthy accounts with realized value that nobody has asked to be a reference yet. For an SE, it scores each open POC's technical risk with the next technical action. For a BDR, it surfaces the warmest prospects (those showing real buying signals) to reach while the signal is fresh. For a founder or GTM operator, it ranks the warmest accounts (including your TAM targets) to work today. For a sales manager, it reads run and outcome aggregates for the reps on the teams you own (never a rep's content) and surfaces who needs a 1:1 first, checks each rep's open pipeline against the number they carry, and builds the commit back up from the deals inside it, so you see which committed money still has no close date, no economic buyer, no agreed next step, or nothing moving on it. A rep whose CRM carries no forecast category is reported as having called nothing, rather than being given a number nobody asked them for. For a CRO, it runs the governed-model hygiene gate over the pipeline and flags the deals whose data quality would undermine the forecast. For a partner manager, it tracks partner certification state and flags the partners missing a required cert or with one expiring soon. For a support lead, it surfaces the accounts whose health has turned, the same read the CSM seat gets, so an open escalation sitting on an account already in trouble is the one you see first, and it clusters your tickets to find the issue that keeps coming back. A cluster only reaches you as a product ask once it has arrived across more than one month and somebody escalated it or rated it bad; it comes with the ticket count, the customer count and the months, so you can take it to product as it stands. These feed the same weekly Coaching card, which points at your top gap and the play to run for it. A missing field is shown as a gap to close, never a number invented. The checks read your 1,000 most recently updated records of each kind, so a very large import is scored on the freshest slice rather than the whole history.
• Recommended for your seat. Settings shows the connectors your persona gets the most from, so a BDR sees its prospecting tools, an SE its demo and RFP tools, and a CSM its success platform, on top of the CRM, email, and calendar core every seat needs. The list is split in two, and the split is the useful part. Connect today is what you can link and use in the next ten minutes. On request is written and not usable yet, and each one says why in plain words: either it needs an approval from the vendor before anyone can use it, or it was written from the vendor's documentation and has not been run against a real account. A count of logos where a third of them do not work is a count that misleads, so the two are separated rather than distinguished by a small badge. Connect the managed CRM, email and calendar right there; the rest connect on the Connect page.
• What your market talked about. The community-monitor and competitor-watch skills can read the list of public Hacker News threads, X posts (only when you saved your own X token in Studio) and company news from the GDELT Project that used your ICP topic keywords, a competitor's name or handle, or one of your accounts, each line labelled with its source, its link and the term it matched. When you connected your own Google Search Console, it also carries the top Google searches that showed your site in the last 28 days, with clicks, impressions and average position, in a block of their own. Everything in it is text written by strangers and is passed to the skill as data, never as instructions. Community monitor suggests which of those threads deserve a reply from you. Competitor watch adds a short section on what the market said about each competitor, and never treats a thread as proof that a competitor is in one of your deals. The list is free to build and reading it adds no connector. Set the keywords and see the list on Studio.
• Do answer engines cite you? (gtmcron, monthly) In the first week of each month we ask up to ten buyer questions built from your confirmed brief (questions about your category, about the problem you solve, and comparisons with the competitors on your brief) of Claude with web search, one search per question. We record which sites each answer cited and whether it named you, and we ask one more question, "what is your company?", to compare your name, one-line offer and category with your brief, the last rescan of your own website and your Brand & Company name. On the 8th the answer-engine-check skill drafts the report: where you were named, the sites and mentions gained and lost since last month, each mismatch, and two or three fixes for each page the findings point at. The draft waits in your queue for review like every other. The answers come from one engine, which the report names; ChatGPT or Perplexity may cite differently. The searches need a Claude key, because web search is a Claude tool. The check runs on your own Claude key when you saved one. Where your plan asks you to bring your own key, a month without one is skipped. Otherwise it runs on the included budget only when that budget has room for the whole run. A skipped month says why.
• Signals from your connectors. Once you connect a doer tool, its activity feeds the free account-signal check automatically. An Outreach reply, a Salesloft meeting, and a Planhat health-score drop each land on the matching account as an unactioned signal, so the deal-risk sweep flags accounts that need a touch without any AI spend. Signals attach to the account by matching the contact's email (or, for Planhat, the company) to what your CRM already synced.

Bring your own key (BYOK)

By default the runtime uses our key with a monthly budget cap priced into your subscription, so it works with no LLM key of your own. In Settings you can instead supply your own Claude, OpenAI or OpenRouter key and choose the model it runs on; runs then meter to you and don't draw our budget. Keys are write-only: once saved, the console only ever shows whether a key is set, never the key itself. Your key powers everything the console runs for you: scheduled hosted runs and the Live coach tools (roleplay, meeting prep, inbound reply) alike. Bedrock support is built but not accepting keys yet; the option unlocks once the transport is validated, and until then the Settings card shows it as coming soon.

Because a saved key is encrypted at rest, it is possible for a stored key to stop being readable, for example after a server-side key rotation. The console checks this rather than assuming a saved key still works: when a key is stored but can no longer be read, the provider shows as key set, unreadable, the LLM pill reads key unreadable, and the Settings card tells you to paste the key again. Scheduled runs keep going on the included tier in the meantime, and each such run records the reason on its own row in the Runs tab, so a run billed to the included tier instead of your key is never silent. Where a deployment requires your own key, that run stops instead and the Runs tab shows the failure with the same reason.

Provider and model (since 12 September 2026). Pick the provider, then the model from that provider's list; the recommended one is preselected, and Custom model id takes an id the list does not carry. With a Claude key each role runs on the model its task was routed to. With an OpenAI or OpenRouter key and no model chosen, each role runs on that provider's model for the same tier (GPT-5 for drafting, GPT-5 mini for light work; on OpenRouter, Claude by way of OpenRouter), so a key saved without a model never sends a Claude model name to OpenAI, which is what used to fail every run. A model you choose applies to every role on the account.

Checked on save. Pressing Save and check key sends one tiny request to the provider with the key and model you chose. If the provider refuses it (a wrong key, a model it does not serve), nothing is saved and the card shows the provider's own message. If the provider cannot be reached, the key is saved and the card says the check did not complete; the next run reports a real error if the key turns out to be wrong. Remove key takes the account back to the included tier.

The scheduled runs and the per-account crons (the morning queue, follow-up extraction, gap questions) all spend the same key. The one call that still needs a Claude key is the contact research behind the morning drafts, which uses Claude's web search; with only an OpenAI or OpenRouter key that one call runs on the included tier and the draft itself runs on yours.

Your knowledge library (grounding)

Every AI draft grounds on your proof points, not general knowledge. The console keeps a per-tenant library of cited notes: case studies, win stories, product facts, each traced to a named source. Before a run drafts, its library is pulled in as a cited-context block and the model is told to cite the source in brackets after each claim and to use only those facts. If the library has nothing on what a draft needs, the run says so plainly rather than inventing a stat, and any number in the draft that lacks a citation is flagged for review.

Your library has two layers. On your first scheduled run the console seeds the seat's built-in craft library (your persona's playbook: value framing, discovery, qualification, and so on), so grounding is never empty on day one. On top of that you add your own proof points in the console's Library tab: title a win, list the facts, and it is cited from then on. The Library tab shows both layers separately, so you can review exactly what grounds your drafts, add or remove your own proof points, and remove any built-in craft you do not want. Personal, medical, and signed-legal sources are refused at ingest. Your library is private to your org and never visible to another tenant.

Not every skill is grounded, and the ones that are not say why rather than staying quiet. A skill that narrates a score the deterministic tier already computed takes no craft input, because no playbook text can change a number that is already decided. Since 2026-08-22 the founder-marketing skills on gtmcron (launch playbooks, the content calendar, messaging experiments, build-in-public, community replies) read a content craft library of their own rather than the deal-running playbooks: writing for an audience who has not raised a hand is a different discipline from running a deal, and grounding a launch plan in objection handling would have produced a citation that pointed somewhere misleading. Two gtmcron skills stay deliberately ungrounded for reasons that are not about a missing library: the investor update reports to a board, and the decision log records a decision already made.

Two layers is the whole list. Until September 2026 a scheduled run also passed its playbook text through a per-team override layer, which could substitute a team's own wording for a built-in playbook file. Its editor was retired long before, so nothing could write an override and the layer changed nothing on any run; it was removed rather than left as a step that reads a database and returns what it was given. Nothing a run is grounded in changes as a result.

The two layers play different roles. The built-in craft grounds how to frame a draft (the method), but a run's proof points, the stats, logos, and quotes it can cite as results, come only from what you uploaded. So a draft never presents a generic playbook line as if it were your customer's outcome; if you have not added a proof point yet, the run says so rather than borrowing one. Extend or replace the seeded craft with your own notes anytime.

A third layer arrives once your organisation has more than one seat: what colleagues have already established. Every run can read the findings other roles' runs recorded about an account, so a customer success run that spotted a renewal risk is something the account executive's run reads rather than works out again. The run sees whether a finding is its own or a colleague's, and it is told to check before deriving an account fact itself. You only ever see what you are entitled to see: a finding stays private to the seat that made it unless the type is one your organisation has already declared shareable, and that decision is made once, in one reviewable place, rather than per run.

Since 2026-08-29 a run also carries your own memory: the open gaps, pending next actions and recent insights the weekly reflection found in your activity, appended after the playbook context. Until then that block reached the in-app assistant and nothing else, so a scheduled run was coached from the playbooks alone. It is about two hundred tokens when a reflection exists and nothing at all when one does not; a rep who has not been reflected yet is drafted for exactly as before.

Connect your CRM (managed)

In Settings you can turn on managed connectors. Link Salesforce, HubSpot, Gmail, or Google Calendar right in the browser; we custody your OAuth token encrypted, poll it on a schedule, and fill your accounts, opps, and contacts server-side. There is nothing to install. The panel shows each provider's connection state and last sync, and a Sync now button pulls your book immediately.

Opt-in and reversible. Token custody is a choice: a security-strict team can leave it off and run its own self-hosted delivery instead (see below). Same as the Settings → Connect page, which carries the same panel.

Connected activity tools feed the free deal-risk checks as account signals: Outreach replies, Salesloft activity, Planhat health changes, Consensus demo engagement, Intercom conversations (including Fin resolutions), and Crossbeam partner overlaps each attach to the matching account so the next scheduled pass can flag it. A signal that cannot be matched to one of your accounts is skipped, never guessed.

When a call ends, a draft is waiting. For an AE seat, a recorded call from Gong, Fathom, Zoom, Google Meet or Teams that matches one of your accounts gets one draft in the AE console's Draft queue. If the transcript shows a buying signal (a price discussed, a scope asked for such as a seat count or a quote, or a start date), the pricing-proposal-builder skill drafts a proposal for that account, with the lines from the call quoted under it and the assumptions listed. If not, you get a short follow-up note built from the next steps the call tool recorded, with no AI call at all. The check is plain pattern matching first; only a call it cannot decide goes to a small model, and only the lines it flagged are sent. Prices still come only from your records: a figure said on the call is listed as something to confirm, and if the draft uses one that your records do not carry, it is replaced with [unconfigured: price]. The run has no tools that write or send, so nothing a buyer says on a call can queue an email or a message. A call gets one draft, ever, even if the call tool sends it twice, and nothing is sent. A call within the last three days that has not matched an account yet gets its draft once the contact is on file. Calls start at most two proposals a day for you, because a proposal spends from the same budget as your scheduled skills; past that a signal call gets the follow-up note, which says so, and you can run the proposal from the console.

Deliver reports to your Google Drive

In Settings, the Deliver reports to Google Drive toggle sends each hosted run's report into your Google Drive "Accounts" folder as a native Google Doc plus a raw .md copy, so you read the report where you already work. A run about a specific account lands in a per-account subfolder (Accounts/<account>); a general run lands in the Accounts root. The same deliveries show on your rollup under "Delivered to your Drive".

Off by default. The toggle needs managed connectors on and the Google Drive connection under Settings, Connections, because delivery uses that connection. The connection asks Google for two kinds of access: reading your files for the business brief, and writing only the files and folders this app creates. An older connection that granted read access only cannot deliver, so it shows as needing reconnect; reconnect once to grant write access. Writing a file into your own Drive folder is delivery, not an outbound send.

Under the toggle, the console names the last three files that landed, with links to each, and says plainly when one did not reach your Drive. A failed delivery means the file write failed, not the run: the report is still in the app and on your Activity page, which is where the full history lives. If the console cannot read that list it says so rather than showing you an empty one, because "nothing has been delivered" and "we could not check" would send you looking in different places.

Advanced: self-hosted delivery (optional)

Most teams use the managed path above. If your security posture requires every credential to stay on your own infrastructure, you can run a small n8n on a machine you control that holds your own OAuth and LLM credentials. Those runs place real Gmail or Outlook drafts and write CRM records on your infrastructure, never ours. Templates live in n8n-templates/ and setup lives in your product docs. Nothing is ever auto-sent.

Keyboard: the Cmd-K palette

From the console (or any signed-in page), press Cmd+K / Ctrl+K or click Search in the top bar to jump to a page, a pending draft by title, or a quick action without touching the mouse. See Home (Today) for the full palette guide.

When a coaching section cannot load

The Coaching tab is assembled from several reads: your goals and pace, the cohort comparison, and your quarter recap. A section whose read fails now says so. It used to disappear, which looks exactly like having no goals set, or no cohort yet.

What runs it

The hosted runtime is server-side cron plus a bounded agent loop. It runs on our servers on a schedule, with nothing to install and no app to keep open. Scheduled AI work runs on your own provider key (bring-your-own-key on every tier), after any platform-funded first-run allowance the deployment turns on.