Hire a crew of AI agents, give them your business, and run the whole thing from one dashboard you own.

You don't rent an assistant that lives on someone else's servers. AI Surfer OS runs on your machine, connected to your AI runtime, working from your business context. Your agents, your keys, your memory, your data — all on your own box.

Think of it as the difference between using a chatbot and running a company. A chatbot answers questions. AI Surfer OS gives you a marketing lead, an operations chief, a finance agent, an engineer, a researcher — a full team that works in parallel, remembers your business, ships real deliverables, and reports back to you.

• You own everything. It's self-hosted — the platform, your agents, your API keys, and your memory all live on your own server. No outside memory service, no vector database you don't control, no vendor holding your data hostage. If you can run a small cloud server, you can own your entire AI operation.
• A real crew, organized like a business. Your agents are grouped into departments — Marketing, Sales, Operations, Finance, Engineering, and more — each with a focused playbook and the right tools. There's a main agent that orchestrates the team and hands work to the right specialist.
• It runs over your runtime, not instead of it. AI Surfer OS connects to a runtime that powers your agents (Hermes, Claude Code, or OpenClaw). It doesn't lock you into one AI provider — it gives you a business layer on top.
• Your crew knows your business. A shared Business Brain every agent reads on every message, plus per-agent memory that builds itself from your conversations, plus reference knowledge you train on your own documents. The more you use it, the more it knows.
• They actually do things. Agents build websites and software, generate images, make and answer real phone calls, run scheduled jobs, connect to your real apps (Gmail, Stripe, your CRM), and — with the Companion — can even use your computer. All from a plain-English message.
• Built around your 10%. The platform is designed to hand you back your time. You define your 10% — the work only you should do — and your crew covers the 90% you want off your plate. A live meter tracks how well that's actually happening.

Founders, operators, agencies, and small teams who want the leverage of a full staff without the payroll — and who want to own their AI stack instead of renting it. If you've ever wished you could clone yourself and delegate the busywork, this is the operating system for that.

In one line: AI Surfer OS is your own AI-run business — a crew of agents you own, working your business, from a dashboard that never leaves your hands.

AI Surfer OS rides on top of an AI runtime — the engine that actually powers your agents. You pick one in first-run setup (and can change it anytime in the connection settings). There are three real options plus a demo:

• Hermes — runs on an always-on server (a VPS). The dashboard talks to it over its API. This is the most common choice for running a business, since it stays live around the clock. If you don't want to wrangle a server by hand, managed hosting like xCloud can stand up Hermes for you in a few minutes.
• Claude Code — the dashboard drives the claude CLI on your machine. Any subagents you've saved in ~/.claude/agents/ automatically show up as your crew. Great if you already live in Claude Code.
• OpenClaw — runs locally; the dashboard connects to its gateway.
• Demo — no runtime at all, just a simulated crew so you can kick the tires before committing.

Most features work the same across all three real runtimes. Where they differ, the docs and the in-app guide call it out.

You log in with a single password (no username). Once you're in, you stay signed in for 90 days, so you're not re-entering it constantly.

• Your dashboard sits behind that password from the very first boot — it's private by default. Nothing to switch on.
• Want more? Optional two-factor adds a phone-code lock (scan a QR with any authenticator app, save the recovery codes). It runs entirely on your own server — no texts, no email, no third party.
• Locked out? From your AI Surfer OS folder, running npm run password prints a fresh one so you can get back in and set your own.

This is the heart of the platform, so do it early. Your 90/10 is your split:

• Your 10% — the lane you're genuinely great at and love, the work to stay in.
• Your 90% — everything else, the work you hand to your crew so you can stay in your 10%.

You can set it two ways:

• Discover my North Star — a warm, guided ~5–10 minute discovery. You dump what's on your mind, it asks a few questions at a time, and quietly builds three things: My Why (your mission), My 10% (your lane), and My 90% (what to delegate). It auto-saves, so you can pause and continue later.
• Type in my own — prefer to write it? Paste your Why / 10% / 90% straight into the three boxes and save.

Why it matters: once set, your North Star is fed to your main agent on every turn — chat, voice, and scheduled jobs — so everything the crew does is steered by it. A live meter then tracks how well your day-to-day actually matches your intent. It's not decoration; it's the compass the whole crew steers by.

Open the agent picker and choose a role or template. This creates a real teammate on your runtime, filed into a department. The ready-made crew ships with executive titles and focused playbooks:

• CFO (Finance), CMO (Marketing), COO (Operations), CTO (Engineering)
• Head of Sales, Head of Product, Head of People, Head of Support
• Creative Director, Legal Counsel, Research & Analysis

Each one arrives pre-loaded with the right skills and tools — Marketing gets content, image, and PDF design; Engineering ships real software out of the box.

Every new hire introduces itself. When it joins, it drops a quick hello in its own voice: what it's great at, and the exact tools it recommends you connect (through the Tool Shed) to make it most powerful. Your sales rep points you to your CRM and email; your researcher to your search tools. Connect what it suggests and it works at full strength.

New installs get a "Set up your system" checklist right on your home base that auto-checks itself as you go. The four things that get your crew grounded and useful:

• Fill in your Business Info (the shared brain every agent reads)
• Set your 90/10
• Create your first agent
• Connect an app (like Gmail or Stripe) so agents can act on real data

Knock those out and your crew stops guessing and starts working from your actual business.

HQ is the first thing you see and the mission-control view of your whole operation, top to bottom:

• Scorecard — key numbers at a glance: agents working now, tasks done today, spend today, and your 90/10 score.
• Morning Brew ☕ — an auto-composed daily standup: what shipped, what failed, what's blocked, and decisions waiting on you. (New users see a setup checklist here until the crew's done its first work.)
• The 90/10 card — your North Star and how well the crew is covering your 90%.
• "Needs you" queue — health alerts and decisions waiting on your call (approvals, silent-failure warnings). Dismissable.
• Business Pulse tiles — real numbers pulled live from your connected tools (Stripe, analytics, email platforms, and more).
• The Swell — a live feed of what the crew just did.

Your crew is shown grouped by department, each agent wearing a live status ring (charging = working, chilling = idle). Your main agent is pinned at the top with a crown.

The Beach is a playful, living view of your crew as a little voxel world — the same team, shown as characters going about their day. It's a fun, at-a-glance status board:

• An agent that's working sits on a recliner with a laptop; idle agents wander, paddle out, or hang by the shore; a benched agent sits under an umbrella.
• Click the boardroom to gather the whole crew for a group deliberation (same as starting a group huddle).
• Click any agent for a quick menu to bench them or send them back to work.
• Big crews show up to 15 characters on the scene, with the rest as tappable chips on a bench row.

Nothing on the Beach is required — the real work happens in Chat, the Lineup, and your schedules. It's just the warm, human face of what your agents are up to.

Agents live in departments — Marketing, Sales, Operations, Finance, Engineering, Content — which you can rename and reorganize to match your org. The ready-made crew carries executive titles so it feels like a real leadership team, and each preset comes with a strong, focused playbook and the right skills pre-loaded:

• CFO, CMO, COO, CTO, Head of Sales / Product / People / Support, Creative Director, Legal Counsel, Research & Analysis.

Presets aren't generic — Legal Counsel can research regulations out of the box; the engineer ships real software; Marketing designs on-brand.

Create an agent: open the agent picker (on HQ or during onboarding) and choose a role. It creates a real teammate on your runtime, filed into a department. Each new hire introduces itself and tells you which tools to connect to make it most powerful.

Edit an agent: in Chat, click the agent's avatar in the header (or open the ⋯ Actions menu → Edit). You get one simple editor with everything:

• Name & Title — what it's called and its role (e.g. "CMO · Marketing"). Edit the title to whatever fits your org; it sticks across restarts.
• Model — which AI model powers this agent. Choices are whatever your runtime is signed in to. Add an OpenRouter key and a wide model menu appears (GPT, Claude, Gemini, Grok, DeepSeek, and more). You can even put one agent on a totally different brain — a local model you run yourself, or a specific cloud model — while everyone else stays on the default. Switching a model is safe: it's health-checked, and if a new model doesn't start cleanly it rolls back automatically.
• Voice — the premium voice this agent speaks with (with a ▶ to hear a sample first).
• Prompt — its personality and instructions (its "soul"). Changes take effect on its very next message — no restart.
• Skills — which skills it specializes in (all, or a focused set).
• Character — which voxel character represents it everywhere.
• Knowledge — the documents this agent has been trained to recall.

Remove an agent: in Chat, open the ⋯ Actions menu and choose Remove. (You can't remove your main agent.)

Your main agent is your executive assistant and orchestrator — its name matches your runtime (Hermes, Claude, or OpenClaw). It runs interviews, keeps the whole operation moving, and is your default point of contact.

The key thing: ask it for something outside its lane and it delegates to the right teammate instead of fumbling it. It knows the whole crew roster, so "have the marketing lead draft this" or "get research on X" routes to the right agent automatically.

You shape a teammate five ways, from a one-second tweak to a permanent rewrite. Use whichever fits:

1. Edit it directly (fastest + most reliable). Open the editor (click the avatar) and change its Prompt, Name, Title, Model, Voice, Skills, or Character. Prompt changes land on its next message.
2. Style & You — tell it how to talk to you, and the facts to never forget. In the editor's You tab (or just by saying it in chat), set three things that are pinned on every message forever, so they can never get crowded out:
   • What to call you — "Zach", not "the boss".
   • How to talk to you — "keep replies short, casual tone".
   • What to always remember — a short, curated list of standing facts and hard rules: "never send anything without my approval", "our business is …". Keep it tight — it's a guardrail list, not a dumping ground. You can set any of it per-agent or apply it to your whole crew. Works for every agent, including your main one.
3. Just talk to it — it remembers facts. Say a fact normally ("our rate is $150", "this client hates jargon") and a quiet once-a-day pass files the durable ones into that agent's own memory.
4. Train it on your documents. Attach real files in the agent's chat and tap "Add to knowledge" — contracts, brand briefs, product docs — and it recalls the actual material later.
5. Fill in your Business Info once. The shared brain every agent reads on every message — what you do, your voice, your customers, your rules. Set it once and the whole crew is grounded.

The theme: tell it once, and it sticks. Whether you edit the prompt, set a pinned rule, drop a fact in chat, or upload a doc — your crew gets more yours the more you use it.

Cards flow left to right through five columns:

Inbox → Assigned → Building → Review → Done

• Inbox — unassigned ideas, not owned yet.
• Assigned — has an owner, but not started.
• Building — actively being worked.
• Review — the agent finished it and it's waiting for your sign-off.
• Done — you approved it.

Agents create and move cards as they work. You can also drag cards yourself, add one with New task, and assign an agent right on the card. The board syncs across all your devices, so it looks the same on your laptop and your phone.

Here's the one thing to understand about AI agents: they're reactive. They do work when messaged, not on their own. So when an agent plans a batch of tasks, the cards land on the board but don't advance until the agent is re-engaged. That's normal — and it's exactly what the Nudge button is for.

• Nudge — on any Assigned or Building card, tap Nudge to tell that agent (in chat) to actually do the work and move the card to Review. Nudging an Assigned card also slides it into Building.
• Nudge all — each column header has this to push everyone in that stage at once. It sends one grouped message per agent, so no spam.
• After you nudge, the card shows "Nudged Xm ago" until the agent speaks, then flips to "Agent replied" — so you can see at a glance which ones answered.

Nothing leaves Review on its own — this is your quality checkpoint:

• Approve → the card moves to Done, and the agent automatically picks up its next queued task.
• Send back → tell them what to fix; it returns to Building for rework.

This is the safety valve of the whole system: your crew can build all day, but nothing ships without your sign-off.

There's an Auto-Nudge toggle in the Lineup toolbar, off by default — leave it off if you like to interject. Flip it on and the crew re-engages its own stalled Assigned/Building cards on a steady cadence (about every 45 minutes, while the dashboard's open), so work keeps moving without you tapping Nudge.

Auto-Nudge never auto-approves Review. Your sign-off always stays manual, no matter what.

Work doesn't get stuck with the wrong person. When one agent hits something outside its lane, it hands the task to the right teammate by name — and that assignee gets dispatched immediately, with the result reporting back into the original chat. You can trigger this yourself too: just tell your main agent "have the designer handle the graphics" and it routes the work.

So the board doesn't pile up over time:

• Mark a project Done and its cards tuck off the board automatically — archived and fully recoverable. Set the project active again to bring them back.
• The Done column shows only the most recent few, with a "show older" toggle for the rest.

Open the Chat page, pick an agent up top, and type. That's it. The agent works from your Business Info, its own memory, and any documents you've trained it on — so it answers in context, not from a blank slate.

• Quick tools live on the + button by the message box: attach a file or image, run Consensus (below), and more.
• Copy anything — every message bubble has a copy icon (next to the speaker icon). The copy icon in the chat header grabs the whole conversation.
• Hear it out loud — each message has a speaker button to read it aloud in that agent's voice.
• On a phone — tap the collapse control in the thread header to tuck away the top bar for more room; the ⋯ menu holds per-thread actions (edit the agent, copy, clear, remove).

Agents don't just chat — from a plain message they can schedule jobs, move board cards, remember facts, build things, send you one-tap questions, and more. You never type special syntax; you just ask.

Facing a decision you really want to get right? The Consensus tool (under the + button) asks the top AI models the same question and hands you back one synthesized answer. Instead of trusting a single model's take on a big call — a pricing decision, a legal-adjacent question, a strategic bet — you get the collective read, reconciled into a single recommendation.

A Huddle is a group deliberation: you put several agents in a room and let them work a problem together, each bringing its own specialty. Your CMO, your Head of Sales, and your CFO can hash out a launch plan in one thread, arguing the trade-offs from their own angles, so you get a rounded answer instead of one narrow view.

• Start one from Chat, or click the boardroom on the Beach to gather the whole crew for an all-hands.
• It's the fastest way to pressure-test an idea across your whole leadership team at once.

If you run Hermes or OpenClaw, your agent can be reached on messaging apps like Telegram and WhatsApp — it's the same agent behind every door. The Channels page is a read-only window into those conversations:

• Each one is clearly source-tagged (a "via Telegram" chip) so it's never confusing where a message came from, and it never gets mixed into your main agent chat.
• A dot marks conversations with new activity.
• It's view-only here — reply from the app itself, or open a fresh thread in Chat. Anything the agent learned on those channels still flows into your Memory, because it's all one brain.

(Claude Code has no built-in messaging channels, so this page doesn't appear there.)

A simple personal checklist on your home base (full list at /todos). It's the frictionless place to capture anything before it slips:

• Jot something down and check it off yourself (with a little celebration).
• Hand it to an agent from the dropdown — pick a face and it dispatches that agent to take it on (asking first if anything's unclear) and promotes it onto the Lineup board.
• Tap an item to add details or context — that gets handed to the agent so it has the specifics.
• Each item shows how long it's been sitting, and the card tracks how many you clear (done today / this week).

Agents can add to this list too, when you ask them to put something on your to-do list. Rule of thumb: lightweight reminders land here; bigger team work belongs on the Lineup board.

You don't have to manage these — the crew handles them — but here's the whole picture so you know what's happening:

1. Business Info (the Brain) — one shared source of truth about your company that every agent reads on every message. Durable stuff: what you do, your voice, your customers, your rules.
2. Shared Team Memory — a short team whiteboard of "what's true right now" that every agent also reads. Moving, current facts: a decision you just made, what a client said, where a file was saved.
3. Each agent's own memory — agents remember on their own. Once a day a quiet background pass reviews recent conversations and pulls the durable facts out of what you said — a rate, a preference, a date — and files them under that agent. These belong to that one agent.
4. Reference knowledge — real documents you train a specific agent on (contracts for Legal, brand briefs for Marketing). It recalls the actual material later.

You never have to say "remember this." Agents capture the durable facts themselves — and remembering can never take an action, it just writes the note.

Business Info is the curated source of truth every agent works from. Train it once; the whole crew is grounded. It only changes in deliberate ways (so it stays clean), and you have five ways to fill it:

• Edit it directly — guided sections plus a free-form brain-dump, then save.
• "Interview me" — your main agent interviews you one question at a time and fills it in.
• "Upload a doc" — it reads a document you upload and fills in the brain.
• "Pull from agent memory" — already trained your agent elsewhere (on Telegram, in a terminal)? This reads what it already knows and pre-fills the brain for you to review. Perfect right after installing over an existing agent.
• "Upload a brand kit" — drop in your brand guide (PDF, image, Canva export) and the crew reads it into your Brand section: exact hex colors, fonts, logo, voice, imagery style. It actually looks at the pages, so even a flattened, image-only brand kit gets read.

Normal conversations do not change Business Info — only those flows do. You can keep several businesses and switch between them.

Personal Info is a separate, smaller card for your own basics (name, email, phone, address). Fill it once and your crew reuses it to complete forms, sign-ups, shipping, and documents on your behalf — no re-typing. Leave anything blank you'd rather not store.

Want an agent to be genuinely good with your material — filling out your contracts, matching your brand, following your product's real docs? Train it on examples.

• How to add it: open the agent's Chat, attach a file (PDF, Word, .txt, or .md), then tap the "Add to knowledge" chip that appears. That pulls the text out, files it under that agent, and confirms.
• From then on, whenever your question relates to that material, the agent automatically pulls the most relevant passages and uses them — matching the format, terms, and structure — without you re-uploading or reminding it.
• It's per-agent. Train your Legal agent on contracts and your Marketing agent on brand briefs; they don't bleed into each other.
• See or remove what an agent knows in the editor's Knowledge tab.
• It's keyless and private — the doc is read and stored on your box; nothing goes to an outside service.

(Attaching a file just for one message still works as always. "Add to knowledge" is the extra step that makes it stick for good.)

You can also import an Obsidian vault — upload a .zip and AI Surfer OS reads the Markdown notes into your main agent's reference knowledge and mirrors them into the Brain graph so you can watch them land.

The Memory page is where you see and prune everything — presented as a brain graph you can explore. Tap the ⤢ expand button to open it full screen: drag nodes, zoom, double-click for a local view, and click any node to read its actual text at the bottom.

• Your agent's existing memory shows up here the moment you connect — no import needed.
• Every save keeps a backup.

For the big pools (an agent's own facts and its reference docs), the crew never dumps everything into every message. It retrieves just the most relevant pieces for the message in front of it. That matching is plain, keyless search that runs right on your box — it leans on the rare, distinctive words (a term like "indemnification" counts for far more than "the") and prefers recent notes.

That's a great fit for real business documents, which are jargon-heavy and consistent — and it means no embeddings, no vector service, and no extra API key are ever required.

It tidies itself, safely. Once a day, a quiet background "memory gardener" collapses duplicate facts and refreshes the graph so things stay clean as they grow. It's built to never lose anything important: tidying only ever moves a redundant copy to the archive (still findable in search, fully recoverable), never deletes, and never removes the last record of a date, price, or number. Nothing you told your crew is ever gone.

Skills are capabilities your crew can use — proven playbooks for doing real work. In the Tool Shed you Add a skill pack to install it for your whole crew, or Update an already-added pack to the latest version (it keeps a backup).

Packs include:

• Vibe Coding — build and ship real software, including stunning websites.
• PDF & Documents — a real design system that makes beautiful, varied PDFs (multiple distinct studio looks, not the same template every time).
• Image Generation, Research, Outreach & Email, Content & SEO, and more.

A few things worth knowing:

• Presets come pre-loaded. When you hire a preset agent (Marketing, Sales, Engineering…), the OS installs the right packs and focuses that agent on them automatically — so they produce great work out of the box.
• Your crew learns on its own. When an agent builds itself a new skill mid-job, a "New skill learned" badge pops up in that agent's chat — you literally watch your crew getting better as they work. Skills are shared across the whole crew: once one learns something, everyone can use it.
• Assigning skills to an agent is a "reach for these first" hint, not a hard limit — it just focuses an agent on its lane.

Does stocking lots of packs slow agents down? Barely. Agents don't load every skill into memory — they see a lightweight catalog (each skill's name plus a one-line description) and only pull a skill's full playbook when they actually use it. So stock the shelf freely; the per-message cost is about one line per skill, not the full playbooks.

The Tool Shed → Apps is where you connect real services so your agents can act in them — Gmail, Calendar, Stripe, your CRM, and hundreds more. This runs through Composio, a connector layer that handles the auth:

1. Create or log in to your Composio account.
2. Copy the setup prompt and paste it into your main agent chat.
3. Open the login link your agent gives you and authorize the apps you want.
4. Return to the Tool Shed and tap Refresh — the connected app bubbles appear.

Once an app is connected, your agents can read and act on it: triage your inbox, check today's revenue, update a calendar, pull CRM contacts. The Business Pulse tiles on your home base also light up with live numbers from whatever you connect.

GoHighLevel (LeadConnector) has a dedicated card: paste a Private Integration token plus your Location ID and your crew can read and act on contacts, opportunities, calendars, and conversations — works the same on white-label GHL domains.

Secrets is your private vault for API keys and site logins. The safest, fastest way to add one is to paste it right into Agent Chat:

• The OS intercepts the secret before the AI or chat history can see it, saves it to your private vault on your server, and gives the crew only the saved connection name. You get a 🔒 "Saved to Secrets" (or "Saved login") receipt.
• Keys are kept server-side, never shown to agents in the clear.

What you can store:

• API keys — things like Tavily (search), Firecrawl (scraping), OpenRouter (extra models). They wire into your agents and reload automatically. Just tell the agent what a key is for ("here's my Firecrawl key") and it titles it accordingly.
• Site logins — a Name · Site URL · Username · Password for any site your crew needs to sign into (an admin panel, a portal with no API). The password is never shown again and never sent to the chat or the AI — your agents get the name, site, and username so they know the login exists and can sign in for you behind the scenes.

The Tool Shed is the visual place to review, rename, or remove anything you saved — never a required detour before chat.

Under the hood, AI Surfer OS speaks MCP (Model Context Protocol), the open standard for connecting AI to tools. Your runtime can wire up MCP servers, and the platform surfaces which servers are connected. For most people this is invisible — you connect apps through the Tool Shed and it just works — but if you run your own MCP servers, they slot into the same crew.

The short version: connect apps in the Tool Shed, drop keys and logins right in chat, and your whole crew can act in your real tools — with every secret staying locked on your own server.

Cron Jobs (also called schedules) are the tasks your agents run on a repeating cadence — a morning digest, a weekly report, an inbox sweep, a research bot. On the Cron Jobs tab you can turn on recommended jobs in a tap, toggle any job on or off, run one now, and see each job's next run and last result.

The easiest way to make one: just ask. Tell any agent in chat "every weekday at 7am, triage my inbox" and it sets up the job for you. It shows up in your list, owned by that agent, where you can toggle, pause, run, or delete it.

A few things that keep it clean and honest:

• Jobs are grouped by the agent that runs them, each agent a collapsible dropdown, so a big crew with lots of jobs doesn't fill the page.
• The agent you chose stays the owner — its report comes back to that agent's chat, not silently to your main agent.
• Notify in chat vs. run silently — each job has a toggle. Notify in chat (the default) drops the full report from each run right into that agent's thread (but stays quiet if a run finds nothing new). Run silently just works in the background; results still flow into your Morning Brew.
• Build the timing exactly how you want — daily, weekly, or monthly, the specific day(s), and the exact time. Plain English by default; the raw cron expression only shows if you deliberately choose Custom cron.
• Recommended jobs are honest about what they need — a suggested Gmail or Stripe job checks that its connection is ready before it's created, and tells you exactly what to connect if something's missing.

AI Surfer OS keeps a steady background pulse so your operation stays alive even when you're not looking at it. This is what powers:

• Scheduled jobs firing on time, whether or not the dashboard is open.
• The 90/10 score re-measuring itself daily, so the trend line stays honest.
• The daily memory gardener tidying and de-duplicating memory in the background.
• Silent-failure alerts — if a scheduled job misses its slot, HQ gets loud about it rather than letting work quietly stop.

You don't configure the heartbeat; it's the always-on backbone that makes automation trustworthy.

The AI Worker Audit helps you decide what your crew should actually do first — and it's genuinely personalized, not a generic list. It reads your whole picture: your 90/10, your full Business Info and Personal Info, Shared Team Memory, your actual crew, your connected tools, and what's on your Lineup. Then it works like a consultant doing an audit: it finds the gaps in your business and designs the AI workflow that closes each one.

Every recommendation is built as Gap → Fix → Impact:

• The gap — the hole it closes, grounded in a real number from your connected tools wherever possible ("340 contacts and 0 in a nurture sequence", "47 unread emails, 12 needing a reply"). It reads the same live figures that power your Business Pulse.
• The fix — the workflow, assigned to a real teammate by name, with a cadence and the tools it uses.
• The impact — the concrete before → after once it's running, and rough time saved.
• See what you'll get — preview a realistic sample of the very first output (the first email, a sample Monday digest) before you commit.

Launch worker does three things at once: creates a real Lineup project assigned to the named teammate, creates the recurring job (already attached and running) if it's scheduled, and drops you into that agent's chat with the kickoff message pre-written — just hit send.

Nothing you generate is thrown away. Every workflow from every run is saved to Suggested jobs with its agent and cadence attached, so the ones you don't launch today stay one tap away.

The Business Pulse is the grid of live numbers on HQ, sorted into Crew, Money, Reach, and Today. Popular apps (Stripe, Gmail, Calendar, and more) light up automatically with a hand-picked number the moment you connect them; any other app you connect shows up as a tile too.

• Hours saved is the headline (in the Crew row) — how much time your crew saved you. Every finished task is valued by how long it would've taken you by hand (a website counts for more than a quick reply), so it's a real, conservative estimate.
• Tap any tile for the detail behind the number, plus a link to open the app itself.
• Set up or customize any tile in plain English — type what you want ("today's new orders", "total unpaid invoices") or tap Suggest for me to pick from what that app can show. Your agent reads the app live and finds the real value; it never makes numbers up.
• Pick your window — a day / week / month switch (or a custom date range) re-pulls the whole grid.
• Connect your money — tie in financial accounts beyond Stripe (Mercury, Brex, Wise) with a read-only key.

On a ChatGPT login, image generation just works. If your runtime is Hermes signed in with ChatGPT, just ask — "make me a cinematic image of a surfer at sunset for my homepage hero" — and your agent generates it and drops it in Assets. No extra key, no separate billing; it runs on your ChatGPT subscription. New installs are wired for this out of the box.

Your crew can produce banners, thumbnails, ad creative, illustrations, logos, and carousels — composed in one pass and saved to Assets.

A couple of honest notes:

• It draws on your ChatGPT plan's usage, so very heavy volume can run into the plan's limits.
• Want a specific image or video model, or higher throughput? Add one in Secrets — OpenRouter for a wide model menu, or Higgsfield for cinematic video. Optional extras, not a requirement.

If your runtime is Hermes signed in with ChatGPT (OpenAI Codex), your agents can build software through the Codex CLI — a coding-tuned agent that reads the repo, edits files, runs it, and iterates. It's the same brain Hermes runs on, but a much better coding loop, so real builds come out higher quality.

• When it's connected, agents automatically route real coding to it — just say "build me a website that does X" and they run it through Codex, then report what changed.
• It's a one-time setup, no terminal, no extra payment — it runs on your existing ChatGPT subscription. You enable device-code authorization in ChatGPT once, connect it in the OS, and approve from your phone.
• Don't have Codex or on a different runtime? No problem — Hermes still codes with its own tools. Codex just makes it noticeably better.

When an agent builds you a website, you don't have to describe every tweak blind. Open the site in Assets and hit ✏️ Edit website to open the Edit Canvas: your live site on the left, a chat with the exact agent that built it on the right. Tell it what to change in plain English and the preview refreshes the moment it's done.

What you can do in the canvas:

• Talk to the builder — "make the hero bigger, add a sign-up button, use my brand colors." It edits the same file in place — no duplicate "v2" asset.
• Attach references — drag, paste, or 📎 in screenshots, a logo, or product shots. Drop in a logo and it wires it right into the page.
• See it on every screen — a desktop / tablet / phone toggle previews at real device widths, so you catch mobile issues before you ship.
• Try it on your real phone — a QR button shows a code; scan it and the page opens right there.
• Versions & restore — every open and every edit auto-saves a restore point. Hit the 🕘 history button to restore any one in a click (and a restore is itself undoable, so you can experiment fearlessly).
• Fix it in one tap — if the preview throws an error, a red "Fix it" chip appears. Tap it and the agent repairs the root cause. You never read a stack trace.

Deploy lets you pick where the site goes live:

• Vercel (recommended) — free, and the link stays live permanently. Best pick for anything real.
• here.now — an instant link, great for a quick share (permanent with a connected account).
• …or push the source to GitHub — sends the code to a private repo you own.
• …or download it — one click saves the code to your computer.

Either way, your agent sends you the live link once it's up, recorded on the asset.

Your crew can build real software, not just websites — it lands in Assets under its own Software tab, with the same live edit canvas ("✏️ Edit app") and iterate-in-place flow.

• Micro-tools — "build me a mortgage calculator / habit tracker / quote generator." The agent ships a single self-contained app that works immediately, right from Assets — no accounts, no hosting. Your data stays on your device.
• Bigger web apps — a real multi-file app shows up as one card (not a pile of files), previews and edits like everything else, and deleting the card cleans up the whole app.
• Use it on your phone — ask for it "as an app I can install" and the agent makes it a PWA: open it, Add to Home Screen, and it behaves like a native app (own icon, works offline). For most needs, this is the mobile app, shipped the same day.
• Apps with login or shared data — the agent builds against a hosted backend (like a free Supabase project) so nothing heavy runs on your server.
• True native apps — the agent can scaffold an Expo (React Native) project for the App Store / Google Play, with the exact steps to preview and build it.

Quality is baked in. Your engineer checks its own work before it ships — it renders the page in a real browser, screenshots it on desktop and mobile, gets a candid "ship it / needs fixes" read on its own build, and fixes anything broken or ugly first. On top of that, every build runs the AI Surfer Design QA pass automatically: comfortable on a phone (big tap targets, no sideways scroll), usable for real people (labeled inputs, readable contrast), and styled on purpose instead of generic AI slop. Builds are dramatically better with Codex connected.

Real API connections, no leaked keys. Sites and tools your crew builds can call real services (Stripe, email, web search, AI images — anything in your Secrets) through your own box, which injects the key server-side. The built app never contains a key, even when deployed to a public host.

Everything your crew makes — images, videos, docs, audio, and live websites — lands in the Assets gallery.

• Tap any asset to open it big: preview, Download, Open, visit its Live site, or delete.
• Each card shows the face of the agent that made it. While a build is in progress, a slim "still being built" strip shows instead of a half-finished file.
• Folders — make folders and organize assets into them (deleting a folder never deletes your assets; they move back to All).
• Search up top, filter by type, and switch between grid and list.

The AI Surfer Companion is a small app you install on your own computer that lets your crew do things on your machine directly — the same actions you'd take with a mouse and keyboard, done for you. It's how an agent goes from "here's the file you asked for" to "I opened the app and did it."

• Getting set up is four steps: download the Companion, install it, pair it to your dashboard, and you're connected. Once paired, your crew can drive the desktop for tasks that live outside a browser.
• It's yours and it's local — the Companion runs on your own computer under your control, so the work happens on your machine, not a stranger's cloud.
• Safety first — like everything in AI Surfer OS, computer use is built to keep you in control: agents act on your behalf, and anything that leaves the building or can't be undone is gated behind your confirmation.

This is the frontier of the platform's "agents can do anything you can, safely" vision — the bridge between your crew and your real desktop.

The Orb (on the Briefing page) is your voice control room: tap it and just talk to your main agent out loud. It reads your standup aloud, answers questions, and pulls real, live info from your connected apps mid-conversation — and it can chain steps in one breath ("check today's revenue and text it to my partner").

It has two modes, switched right under the Orb:

• Listen-only (the default, 🛡️) — the Orb talks and looks things up, but won't take any action. Safe to explore.
• Full Send (⚡) — lets the Orb actually do things: create tasks, update your board, hand jobs to your crew, send emails, place calls, build and deploy. It always confirms before anything leaves the building and won't do bulk deletes. Flip back anytime.

A few of the niceties:

• Clap to wake — double-clap to summon the Orb hands-free.
• If it doesn't already know something, it sends an agent to go get it rather than guessing.
• Recap — end a real conversation (~10 seconds or more) and the Orb drops one short, friendly recap into your work chat.
• It remembers and learns — when a call ends, the important things you said (preferences, decisions, facts) are saved into your agent's permanent memory (you'll see a "🧠 Remembered: …" receipt). Every new call starts already knowing your saved memories and your recent recaps, so "like we talked about last time" just works.
• Hand work to a specific agent by name — "have Ron write the campaign" (in Full Send) sends the job to Ron, not your main agent.
• Speak your language — it defaults to English but follows you if you clearly speak another language; pick your own or turn on Auto-detect for fully multilingual calls.

(Each agent can also have its own voice, set in its editor. The Orb needs a voice key set; if something's missing it tells you.)

Your agents can make and answer real phone calls — sounding like themselves, knowing everything in your Business Info. There's no separate phone bot to build; the OS creates the call assistant from that agent's own training automatically.

Setup is one-time: save your calling key in Secrets, then in any agent's chat turn calling on, pick a number, and pick a voice.

• Outbound — just ask: "call Mrs. Henderson and confirm tomorrow's 9am." You get a one-tap confirm before anything dials (nothing ever calls without your OK), and a 📞 call report lands in chat when it wraps — outcome, summary, and a transcript your agent can quote from.
• Inbound — your 24/7 receptionist — the agent answers calls on its number around the clock (handled in the cloud, so calls are answered even while your server updates or reboots). Point your company number at it by forwarding calls; most businesses forward on-no-answer or after-hours, so the AI catches exactly the calls that used to hit voicemail.
• Leads don't fall through — reply "log this lead" and the agent saves the caller and job to your connected CRM and puts it on the board. Or flip Auto-log leads and every finished call (inbound and outbound) files itself.
• Teach it how to answer by just telling it (greeting, hours, when to take a message) — it saves your preferences and gets better every call.

If you run Hermes or OpenClaw, your agent can be reached on messaging apps like Telegram and WhatsApp — the same agent behind every door. So you can message your crew from your phone the way you'd text a coworker, and the Channels page gives you a read-only window into all those conversations, clearly source-tagged and kept out of your main chat.

Slack (two-way) goes further: let a teammate or customer message a Slack channel and have an agent answer right there. You connect Slack, pick the answering agent, and route specific channels to specific agents. Replies in outside channels run in a hardened, read-only mode, so an agent can answer without taking actions or leaking private info.

Between the Companion, the Orb, real phone calls, and messaging channels, your crew works the way a real team does — on your computer, on the phone, and in the apps you already use.

A runtime is the AI engine your crew runs on. AI Surfer OS doesn't replace it — it gives you one clean place to run your whole crew like a business on top of it. You pick your runtime in first-run setup and can change it anytime in the connection settings. There are three real choices:

• Hermes — runs on an always-on server (a VPS); the dashboard talks to it over its API. This is the go-to for running a real business, since it stays live 24/7 — scheduled jobs fire, phone calls get answered, and your crew keeps working whether or not your laptop is open.
• Claude Code — the dashboard drives the claude CLI on your own machine. Any subagents you've saved in ~/.claude/agents/ automatically appear as your crew. Ideal if you already work in Claude Code.
• OpenClaw — runs locally; the dashboard connects to its gateway.

Most features work the same across all three. Where they differ (for example, messaging Channels are a Hermes/OpenClaw capability), the platform tells you.

This is the core promise: AI Surfer OS is self-hosted. The platform, your agents, your API keys, your memory, and every asset your crew makes all live on your infrastructure. No outside memory service, no vector database you don't control, no vendor holding your data.

Your runtime needs to live somewhere always-on. Two paths:

• Run your own server — a small cloud server (VPS) is plenty. You point AI Surfer OS at it in the connection settings and you're off.
• Use managed hosting — if you'd rather not wrangle a server by hand, a managed host like xCloud can stand up a Hermes agent for you in a few minutes: name your server, pick a plan, paste an AI provider key (Anthropic, OpenAI, OpenRouter, or Google Gemini), and optionally connect Telegram. Once it's live, come back to AI Surfer OS and point it there.

Either way, you hold the keys — literally. Your provider API keys and all your data stay on your box.

Your runtime signs in to an AI provider, and that's what powers the crew by default. You're not locked in:

• On a ChatGPT (OpenAI Codex) login, image generation and high-quality coding come built in with no extra billing — they ride your existing subscription.
• Add an OpenRouter key and a wide model menu opens up (GPT, Claude, Gemini, Grok, DeepSeek, and more) — you can even put a single agent on a different brain (including a local model you run yourself) while the rest of the crew stays on the default.
• Switching a model is safe: it's health-checked, and if a new model doesn't start cleanly it rolls back automatically.

When a new version is available, a "There is a new update available" banner drops down at the top of any page, and the Software updates card shows it — tap either to update. A progress bar runs while the app rebuilds and restarts itself, showing your current version moving to the new one (like v1.1.426 → v1.1.427). You never need a terminal for normal use.

• You stay in control. Agents can guide you to the update screen, but they don't install updates from a chat message — you tap Update now yourself, so updates stay visible and deliberate.
• License status shows as a small pill (active / inactive). Your app keeps running fully either way; an expired license just stops receiving new updates until you reactivate.

The platform is built so your crew can do a lot without you ever losing the wheel:

• Kill switch (top bar, every page) — bench one agent, or STOP ALL at once.
• Review gate — nothing on the Lineup ships without your Approve. Voice and agents confirm before anything leaves the building (emails, calls) and won't do bulk deletes.
• Silent-failure alerts — missed scheduled jobs get loud on HQ, so work never quietly stops.
• Web push — phone alerts on failures, plus one daily briefing ping at your chosen hour (no spam by design).
• Spend cap — an optional daily-spend heads-up that warns at 80% and when you cross it, but never stops the crew (and stays quiet on subscription runtimes with no metered cost).
• Low-disk guard — since your crew saves real work to your server's disk, HQ warns you if free space runs low, before saves can start failing.
• Private by default — your dashboard sits behind one password (stored as a hash, never plain text), agents and keys stay on your server, changing your password logs out other devices, and optional two-factor adds a phone-code lock. Nothing to switch on — it's locked down from the first boot.

Because you own it, you can back it all up. Two buttons in the settings:

• Download backup — a small JSON of everything that makes the box yours: Business Info brain, memory, crew, cron jobs, board, chat history, and settings.
• Download everything (+ media) — that same JSON plus one archive of every asset your crew made (photos, videos, docs, websites).

Keep the two files together — that pair is your full backup. API keys and logins are never in a backup (reconnect those once after restoring), and secret files are left out of the media archive too.

The whole point: your AI operation is yours. Your server, your keys, your data, your crew — portable, private, and always under your control.

AI Surfer OS rides on top of your agent runtime (Hermes, Claude Code, or OpenClaw). That runtime has to live somewhere that's always on, usually a small cloud server (a VPS). xCloud is managed hosting that stands up a Hermes agent for you, so you don't have to set up and babysit a server by hand.

You don't strictly need xCloud, any always-on server works, but it's the smoothest path if you'd rather not manage infrastructure yourself.

The short version of the Hermes deploy flow:

• Start from xCloud's Hermes deploy page, name your server, and pick a plan.
• Finish signup and payment if you're new.
• Paste an AI provider API key (Anthropic, OpenAI, OpenRouter, or Google Gemini) when it asks.
• Optionally connect Telegram so you can message your agent from your phone (you can also do this later from the dashboard).
• Manage everything from the xCloud dashboard once it's live.

Full walkthrough: xcloud.host/docs/how-to-deploy-hermes-agent-with-xcloud

Once the runtime is running on your server, come back to AI Surfer OS and point it there in Settings → Connection. This is about hosting the runtime itself, not something you configure inside the dashboard.

If you set Hermes up with ChatGPT (OpenAI Codex) during deploy, do the one-time Codex connect afterward (Settings → Agent Runtime → Set up Codex) so your agents can code with the Codex CLI. You approve it from your phone or laptop; the server never needs a browser.

When something on your server needs a nudge, the xCloud Command Center is a guided troubleshooter built for the xCloud Web Terminal. Instead of hunting for the right command, you find your issue and it hands you one safe command to run.

How it works:

• Find the issue that matches what you're seeing, organized by area: Dashboard, Access, Runtime, and Diagnostics.
• Click Copy to grab the exact command.
• Open the xCloud Web Terminal from your server (Open Terminal, top right), paste it, and run it.
• If the first fix doesn't do it, the tool walks you to the next exact step, so you're never guessing.

Open the command center: xcloud-command-center.vercel.app

Your agents don't just chat — they act on your business from a plain-English message. No special syntax; you just ask. They can:

• Schedule recurring jobs — "every weekday at 7, triage my inbox" becomes a real cron owned by that agent.
• Run the board — create and move task cards, lay a multi-step build into phases, and deliver finished work with a link or file.
• Build and host websites and software — a landing page, a micro-tool, a full web app, even a native mobile app — and register the live URL.
• Generate images — banners, thumbnails, ad creative, logos, carousels — saved to Assets.
• Make and answer real phone calls — outbound (with your confirm) and inbound 24/7.
• Fill a form for you — open a real page and fill it from your Personal Info, then screenshot it for review (it never submits on its own).
• Use your computer — with the AI Surfer Companion installed, act directly on your desktop.
• Remember things — save a company fact to Business Info, pin a "what's true now" note to Team Memory, or file a personal detail.
• Make a PDF, send a voice message, ask you a one-tap question, hire a teammate, set the day's focus — and more.

Not sure if the crew can do something? Just ask — the agent can read the guide to answer accurately.

Five ways, from a quick tweak to a permanent change:

1. Edit it directly — click the agent's avatar in chat to open its editor (Prompt, Name, Title, Model, Voice, Skills, Character). Prompt changes land on its next message.
2. Style & You — set how it talks to you, what to call you, and a short list of facts to always remember (pinned on every message so they never get crowded out). Set it in the editor or just by saying it in chat.
3. Just talk to it — say a fact normally and a daily pass files the durable ones into that agent's memory.
4. Train it on documents — attach a file in chat and tap "Add to knowledge."
5. Fill in Business Info — the shared brain every agent reads.

• Talk to an agent? → Chat page, pick the agent up top, type.
• Create my first agent? → The agent picker on HQ, or the onboarding "Create your first agent" step.
• Change an agent's personality, model, voice, or skills? → Chat → click the agent's avatar → edit it all in one place.
• Teach the crew about my business? → Memory → Business Info: edit it directly, or tap "Interview me" / "Upload a doc."
• Make one agent remember a document? → Open that agent's chat, attach the file, tap "Add to knowledge." It'll recall the material from then on.
• Set my North Star / 90-10? → The 90-10 card on HQ: "Discover my North Star" for the guided flow, or "Type in my own."
• Push a stuck task forward? → On the Lineup, tap Nudge on the card (or Nudge all in a column header) to re-engage the agent.
• Schedule a recurring job? → Just ask an agent in chat, or use the Cron Jobs tab. It stays owned by that agent.
• Give an agent an API key or website login? → Paste it right into chat and say what it's for. The OS saves it safely before the AI or chat log sees it, and shows a "Saved to Secrets" receipt. Manage it later in the Tool Shed.
• Connect an app (Gmail, Stripe, my CRM)? → Tool Shed → Apps, through Composio.
• Change a website I already made? → Open it in Assets and hit ✏️ Edit website for the Edit Canvas (live preview + chat with the builder), or just ask the agent. Either way it edits the same asset in place.
• Talk to my crew by voice? → The Briefing page Orb. Listen-only just talks and looks things up; flip Full Send to let it act (it confirms first).
• See what things cost? → The Usage / Spend page.
• Stop everything fast? → The kill switch in the top bar (STOP ALL).
• Get alerts on my phone? → Turn on web push in Settings.
• Have my agents reply in another language? → Settings → Your team → Language.
• Organize my assets? → Assets page → the folder icon, then "Add to folder" on any asset.

No. Memory and reference knowledge run on your own box with plain, keyless search — no embeddings, no vector service, no extra API key. It leans on the rare, distinctive words in your documents (which is a great fit for jargon-heavy business material) and prefers recent notes. It's deliberately keyless by design.

Yes, from the first boot. Your dashboard sits behind one password (stored as a hash, never plain text). Your agents, keys, and data stay on your own server. Changing your password logs out other devices, and optional two-factor adds a phone-code lock with recovery codes — all running on your own server, no third party. Nothing to switch on; it's private by default.

That depends on your runtime and AI provider, not the platform. On a ChatGPT or Claude subscription, there's no metered API cost — the dollars read $0 and you're working off your existing plan (image generation and coding included on ChatGPT). On an API-key/metered setup, the Usage page tracks per-agent spend with a 14-day trend and an optional daily heads-up. Either way, the Usage page also shows "What your crew is worth" — the hours saved and value produced, right next to what it cost.

• Stuck agent (a spinning status dot, a "sign-in expired" message)? → Settings → Connection → Reconnect your agent, and hit Restart agent first — it fixes most disconnects. The next step adapts to whatever model your agent runs on.
• "Image generation unavailable"? → One tap in Settings → Agent Runtime → Reconnect → "Turn on image generation" wires it up. (You never need a special image key on a ChatGPT login.)
• A website preview throws an error? → Tap the red "Fix it" chip in the Edit Canvas and the agent repairs the root cause.

• Update → a banner appears when a new version is ready; tap Update now and it rebuilds and restarts itself. No terminal needed.
• Back up → Settings → Security & system → Backup & Restore. "Download backup" saves a JSON of everything that makes the box yours; "Download everything (+ media)" adds an archive of every asset. Keep the pair together.
• Move servers → deploy your runtime on the new box, point AI Surfer OS at it in Settings → Connection, and restore your backup. (Reconnect API keys and logins once — they're never stored in a backup.)