Today we're launching OM2. Your AI re-reads your entire company from scratch every time you ask it something. It’s why more than 50% of your token bill isn't in the answer, but in the search for it. OM2 gives your AI a permanent memory of your company. 🧠

Your CRM. As a command center. Live account health, generated risk scores, and one-click AI actions, on top of your HubSpot, Salesforce, or warehouse. Most CRM views answer one question: what happened? They tell you about the last meeting, the last email, the last update someone remembered to log. This view answers a different question: what needs to happen next? Horizon Labs: health 30/100, risk 33/100. Nova Systems: health 29/100, risk 42/100. Apex Partners: health 58/100, risk 47/100. Every account scored. Every score generated from what's actually in the system: meeting recency, engagement patterns, support volume, contract signals. Not a manual assessment someone filled in last quarter. A live read that updates as the data changes. On the right: suggested actions. Summarize my accounts. Prep for upcoming meetings. Ask about Horizon Labs. The AI doesn't just show you the problem. It shows you what to do about it. Your CRM already has the data. This is what it looks like when the data does the work. What's the account on your list right now that you know is at risk but haven't had time to look at?

Describe a model. Get a working spreadsheet, formulas and all. Rows, formulas, and real structure from the first prompt. The standard way to build a revenue model: open a blank sheet, set up your columns, write the formulas by hand, check the references, fix the broken ones, add the accounts, pull the numbers from wherever they live. An hour minimum, usually more, and that's before anyone asks you to change the growth assumptions. This is the Q2 Revenue Model built in Coworker. Seven accounts. ARR, Q1 actuals, growth rate, Q2 forecast. Formula bar showing =C2*(1+D2). Every cell calculated. Total row at the bottom: $2,283,000 ARR, $570,750 Q1 actuals, $638,728 Q2 forecast. Not a table someone typed in. A working spreadsheet with real formula logic. The kind that recalculates when you change a number. The model that used to take an hour to set up from scratch now takes a prompt. The formulas are there from the start. The structure is right the first time. The only thing left to do is the thinking the model was built to support. What spreadsheet does your team rebuild from scratch every quarter?

Ask for the view. Get a dashboard connected to the real numbers. Filters, live sync, and every metric you asked for. The standard customer health review: someone builds a spreadsheet, pulls a manual export from the CRM, formats it, shares it in Slack, and by Thursday the numbers are already stale. The meeting happens anyway. The decisions get made on data that's a week old. This is the Customer Health Dashboard built in Coworker. One ask. Live data. Active accounts: 1,243. Up 12%. At-risk: 38, down 9%. Average health score: 74, up 6 points. Churn this quarter: 3, down 2 versus prior. Engagement up 23% week over week, with Sunday the strongest day of the week. Not a snapshot. Not an export. A dashboard that updates when the underlying data updates. The metrics that used to require a recurring ops task now require a prompt. The view that used to live in one person's spreadsheet now lives anywhere you want it. What customer metric does your team not have a clean view of right now?

Watch it work, not just wait for it. Every run logged: what came in, what went out, what needed a human. Most automation is a black box. You set it up, it runs, you hope it's working. When something goes wrong you find out from a customer, not from the agent. This is the Infosec Questionnaire Responder. Live. 12 runs this week. 92% auto-resolved. 9:14 AM: 18 questions drafted, cited SOC 2 report. Answered. 7:02 AM: 2 low-confidence answers flagged for review. Needs review. Yesterday, 4:41 PM: 6 questions drafted, approved in 12 minutes. Answered. Yesterday, 11:20 AM: contract terms question routed to legal. Escalated. The agent knows what it knows, flags what it doesn't, and routes what it can't handle. Every decision logged. Every handoff visible. Nothing disappears into a queue. 92% handled without a human. 8% escalated to exactly the right person. That's not a black box. That's a system you can actually trust. What's the workflow at your company where nobody knows if it's actually running until something breaks?

If you can describe it, you can build it. Connected to all your data. Every artifact is live code, shareable instantly. The old workflow: define the spec, queue a sprint, wait for the build, review, iterate, ship. Weeks for anything interactive. Days for anything with data. The new one: describe what you need. Get a live artifact. A game. A calculator. A dashboard. A spreadsheet model. An internal tool. An agent. Every one of these shipped from a single sentence. Not a mockup. Not a wireframe. Real, working code. Connected to real data where it matters. Snake. Mortgage calculator. Account health dashboard. Onboarding flow. Runway model. Support triage agent. The common thread isn't the complexity. It's that none of them required a developer. None of them required a design handoff. None of them required waiting. What would your team build first if the build step disappeared?

Describe the meeting. Get a boardroom deck, every number sourced. Pulled from your CRM, billing, and transcripts. Not a blank template. The standard way to build a board update: someone owns the deck, someone pulls the CRM export, someone else digs through the billing system, a fourth person finds the transcript notes. Then you reconcile all of it into slides the night before the meeting. Two hours minimum. Usually more. This is the Q2 Board Update built in Coworker. Five slides. Pipeline coverage at 3.2x with $13.4M in active opportunities across 8 stages. Discovery through Closed Won, every count and dollar value sourced directly. Key metrics alongside it: $300M ARR, 16,000 customers, $19K ARR per customer, 69% YoY growth. Not numbers someone typed in. Numbers pulled from the systems that track them. The insight at the bottom: strong performance in late-stage deals, focus needed on accelerating early-stage conversion. That came from the data too, not from someone's gut read in a pre-call. The deck that used to take a team half a day now takes a prompt. What's the one slide your board always asks about that takes the longest to prepare?

Claude Opus 5 is now available in Coworker. Select it under Anthropic in the model dropdown. Anthropic's most capable model to date. Stronger reasoning, better at complex multi-step tasks, sharper on ambiguous instructions. The kind of model you reach for when the stakes are higher and the answer actually matters. Coworker routes intelligently across models by default. But when you want to drive, now you can. Pick the model. Run the task. See the difference yourself.

Describe the job. Get an agent that runs it on its own. Triggers, tools, and guardrails, wired to your stack. Not a chatbot. Most automation projects stall at the same place: the handoff between what someone described and what an engineer had to build. You write a spec. Someone queues it. Months later you have a workflow that half-works and nobody remembers why it was scoped that way. This is different. You describe what you need in plain language. Coworker builds the agent out for you. Triggers, data sources, permissions, the review step before anything gets submitted. All of it configured from the description. What you're looking at: an Infosec Questionnaire Responder. Slack-triggered. Watches #infosec-responses#. When a vendor questionnaire drops in, it reads your security wiki in Notion, your policies in Confluence, your SOC 2 evidence in Google Drive. Drafts a response. Routes it for review. Submits it. Every time. Without a ticket. Without someone hunting for the policy doc. Without waiting for the one person who knows where everything lives. The setup takes minutes. The agent runs indefinitely. And you didn't write a line of config to get there. What's the first workflow your team would hand off?

Watch the pipeline move, stage by stage. Every deal, every stage, refreshed the moment your CRM changes. Most pipeline reviews start with the same question: is this up to date? Then someone pulls the CRM, someone else screenshots a table, a third person pastes it into a slide. By the time it's in the meeting it's already wrong. This one works differently. Tell Coworker what you want to see. It pulls from your CRM directly and builds a live dashboard. Q3 Pipeline. Every stage. Deal count per stage. Top open deals, their value, their close date. Synced two minutes ago. Discovery: 24. Qualified: 16. Proposal: 9. Negotiation: 5. Closed won: 3. All visible at once. No export. No formatting. No waiting for the one person who owns the spreadsheet. The pipeline review that used to start with thirty minutes of prep now starts with the actual conversation. What does your team spend the most time building before the meeting even starts?

If you can describe it, you can build it. Connected to all your data. Every artifact is live code, shareable instantly. The old workflow: define the spec, queue a sprint, wait for the build, review, iterate, ship. Weeks for anything interactive. Days for anything with data. The new one: describe what you need. Get a live artifact. A game. A calculator. A dashboard. A spreadsheet model. An internal tool. An agent. Every one of these shipped from a single sentence. Not a mockup. Not a wireframe. Real, working code. Connected to real data where it matters. Snake. Mortgage calculator. Account health dashboard. Onboarding flow. Runway model. Support triage agent. The common thread isn't the complexity. It's that none of them required a developer. None of them required a design handoff. None of them required waiting. What would your team build first if the build step disappeared?

Ask for the brief. Get an export-ready PDF, formatted to share. Proper headers, sections, layout. Ready to attach or print. Not a rough draft you need to clean up before sending. The follow-up brief used to mean an hour of formatting after the meeting. Pull the notes, structure them, add the stats, make it look like something you'd send to a client. Then someone asks for a change and you do it again. Now you describe what the meeting covered. Coworker builds the PDF. Four pages. Executive summary, what you aligned on, the numbers, the plan. Branded and ready to go. 80% lower spend vs. frontier API rates. 6 days from kickoff to first live workflow. 50+ connectors ready on day one. The brief that used to take an hour ships in under two minutes. The next meeting starts with a document, not a blank page. What's the last brief your team sent that took longer to format than to write?