Drafts the reply, proposes a time, prepares the post, writes the report.
An AI front office that drafts the routine work of a business day, for any kind of company. What goes out follows the approval level you set: ask you first, or act on its own for routine tasks.
To [email protected]. Re: New client this quarter?
Hi Maria, yes, we are taking on new clients this quarter. Monthly bookkeeping and VAT filing for a small café starts at ₱4,500. I can meet you Thursday at 17:30 or Saturday at 10:00. Which suits you?
ApproveEditSkip
A day with your AI team.
Five moments from one working day, in order. Each shows the draft it produced and who approves it. Everything here is Live or Pilot. The business is invented: a small accounting practice. The same steps work for a clinic, a shop, a studio or a property office.
-
A customer writes before you are up.
From Maria SantosDo you have room for new clients this quarter? I need help with bookkeeping and VAT for my café. Can we meet this week?
Draft reply, ready for youHi Maria, yes, we are taking on new clients. Monthly bookkeeping and VAT filing for a small café starts at ₱4,500. I can meet you Thursday at 17:30 or Saturday at 10:00. Which suits you?
Who approves: you, from your phone. In Ask me first mode nothing is sent until you tap Approve. Live in the demo; your own mailbox is set up for you
-
The meeting is booked.
Calendar, after Maria picks SaturdayBooked: Saturday 10:00, first meeting with Maria Santos, 45 minutes. Checked against your calendar, no clash. Invite sent to Maria and to you.
Who approves: you, or no one if you chose Act and tell me. Either way it is written to the audit log. Live in the demo, on the workspace calendar; your own calendar is planned
-
The post is drafted for three networks.
FacebookThe quarterly VAT deadline is on the 25th. Send us your receipts by the 20th and we file on time. Message us to set up a first meeting.
InstagramVAT due on the 25th. Receipts by the 20th, we handle the rest. Link in bio.
LinkedInFor small businesses in Makati: the quarterly VAT deadline is the 25th. We prepare and file for cafés, shops and studios, and keep the books in order all year.
Who approves: you, one card per network. In the demo they publish to a sandbox feed. On your own pages, your team posts them today. Drafting is Live
-
Customers who went quiet get a follow-up.
Draft follow-up to Daniel ReyesHi Daniel, the filing deadline is next week and we still need your March receipts. Shall I book a drop-off for Thursday? Reply STOP to opt out of messages.
Who approves: the staff member on duty. One reminder at most, sent between 08:00 and 21:00, never to a number that replied STOP. Pilot, set up per client
-
The daily report lands.
Daily reportToday: 10 messages in the inbox, 10 replies drafted, 2 meetings proposed, 3 posts drafted. Waiting on you: 15 approvals.
Who approves: nobody. It is a read-only summary for the owner. Pilot; the weekly version is Live in the demo
The business, names and figures are invented.
Everything it can do, and how ready each part is.
88 capabilities in 14 areas. Filter by area or by status. Most roadmap items carry a target quarter, which is a plan and not a promise.
- Live 39
- Works today and you can see it: in the demo workspace, the hosted desktop, the chat on this site, or running for HousingInteractive.
- Pilot 23
- Works, but we set it up for each client by hand. Done for HousingInteractive, running as a prototype or proven in our own agent software, not self-serve yet.
- Roadmap 26
- Designed or planned, not working yet. Most carry a target quarter, which is a plan and not a promise.
Showing all 88 capabilities.
Communication and inbox 8
Read customer emails and draft replies
Reads new customer emails, looks up the product or service record, drafts a reply with the real price and availability, and queues it for approval. Live in the demo; your own mailbox is set up for you.
Email on your own domain
Sends replies and campaigns from your domain through Amazon SES or Resend, with bounce and opt-out handling.
WhatsApp and Viber checks and drafted replies
Checks the chats every 10 minutes, drafts replies in the business's tone, and sends them under the written rules.
Route each inquiry to the right person
Assigns new inquiries to the staff member who owns the account and reassigns them when nobody answers.
Answer visitors on your website
A chat on the website answers questions from your own facts and collects a booking request. It never confirms a booking.
Official WhatsApp, Messenger and Viber connectors
Reads and answers messages through each network's own business API instead of a browser tab.
Shared team mailbox in the portal
Shows incoming mail next to the inquiry it belongs to, drafts the answer and waits for approval.
Response clock with escalation
Starts a clock on every inquiry and escalates to a second person after one and three hours.
Scheduling 5
Propose meeting times from a free calendar
Reads the workspace calendar and offers slots that are actually free, on weekdays in business hours.
Book the meeting
Adds the meeting to the workspace calendar after the customer picks a time, and refuses a clash.
Timed and event-triggered routines
Runs a task on a schedule, or when a file, commit, webhook or connector event arrives, with a cap per hour.
A working-day schedule of recurring jobs
Runs the daily jobs on a clock: chat checks every 10 minutes, an availability report every 15, a 06:30 report, a 14:45 reminder.
Your own Google or Outlook calendar
Connects to your real calendar by sign-in instead of the workspace calendar.
Sales and outreach 7
Invite partners and suppliers by message
Sends a first invitation and at most one reminder, 08:00 to 21:00 only, with an opt-out line on every message.
Daily availability check with partners
Contacts a set list of partners each day to check whether what they offer is still available. For HousingInteractive that is about 40 a day.
Match waiting customers to new stock
Every morning matches people who are still looking to what has just come in, and drafts the message.
Follow up customers who went quiet
Writes a short, timely follow-up for enquiries and new sign-ups that stopped replying, ready for approval.
Opt-out and send-window rules
Honours STOP replies on every channel and only messages people inside set hours.
Email campaigns to a list
Plans and sends campaign email to your sign-ups and contacts through a campaign tool.
Paid ads management
Drafts ad copy and manages budgets on Meta and Google Ads with spend limits.
Marketing and content 5
Posts written per network
Writes a separate post for each network from the real details of the product or service, never inventing facts.
Rewrite product or service descriptions
Replaces a weak description with better copy that keeps every fact accurate.
Word, Excel, PowerPoint and PDF files
Creates and edits business documents, slides and spreadsheets and reads PDFs.
Nightly photo and watermark screening
Checks catalogue photos overnight for another company's watermark and flags the item to be taken offline.
Weekly social calendar and promo videos
Plans the week's posts and promo videos from what you have coming up.
Social media 4
Publish approved posts to a sandbox feed
Publishes approved Facebook, Instagram and LinkedIn posts to a sandbox feed you can see in the demo. Publishing to your real pages is on the roadmap below.
Publish to your real Facebook and Instagram pages
Connects your pages by sign-in and publishes approved posts through Meta's API.
LinkedIn company page posting
Publishes approved posts to a company page through LinkedIn's API.
Read messages from social inboxes
Reads Messenger and Instagram direct messages and drafts replies.
Voice 3
Dictation and spoken replies
Speak a task instead of typing it, and hear replies read aloud, hands-free if you like.
Voice receptionist on a phone line
Answers the phone, books meetings and hands off to a person when it should.
Text messages in the Philippines
Sends and receives SMS through a local provider.
Analytics and reporting 5
Weekly report in plain language
Summarises enquiries, response time and reach, with three charts.
Read Google Analytics and explain it
Reads traffic and channel figures from Google Analytics and turns them into a plan.
Daily brief for the owner
Gives a daily summary of users, inquiries and sales, and a numbered pending list at 06:30.
Cost of every task, shown as it runs
Shows what each run cost and the wallet balance after it.
Status page for the team
A page the owners and team lead read to see what is done, in progress and waiting.
Websites, deploys and dev ops 5
Build and ship website changes through a reviewed deploy
Builds changes, runs the test suite and deploys to the server only after approval.
Guarded database changes with an undo
Runs a change only after taking a backup table, rehearsing it and writing an undo script.
Coding agent with git worktrees, review and checks
Edits code in isolated branches, reviews diffs, opens pull requests and fixes failing checks, asking before each step.
Read-only views of Supabase and AWS
Looks at tables and servers without being able to change them.
Jira tickets from staff
Reads new tickets and replies with status.
Data, documents and knowledge 3
Answer from your own records
Looks up products or services, prices and details when answering a customer.
Catalogue clean-up at scale
Finds duplicate records, prepares merges and builds the exports other platforms ask for.
Remembers what you approve
Keeps approved facts and preferences between sessions, kept in your hosted desktop.
Finance and admin 5
Prepaid wallet with every charge on a ledger
Pays for AI use from a balance with an append-only ledger, in test-mode payments today.
Step down to the 4B model when the wallet runs low
When the wallet cannot cover a larger model, a run drops to the untuned 4B model on our hardware, charged at the 4B’s price, instead of failing or overspending.
Live card payments and invoices
Takes real subscription payments and issues invoices in USD or PHP.
Spending limits you set
Set a low-balance alert, a monthly amount that warns at 50%, 80% and 100%, a monthly limit for each person and a budget for each task. A person at their limit, or a task over its budget, stops before the next step. Each plan also has a monthly allowance and a hard cap on the cost of the hosted open-weight model; at the cap, work runs on the 4B model on our hardware.
Templates for clinics, law and accounting firms
Ready-made agents, rules and approvals for these industries.
Control, approvals and audit 12
Draft only, ask first, or act and tell me
You choose how much the AI may do alone. Every action follows that choice.
Approval queue: approve, edit or skip
Every outgoing action waits as a card you can change before it goes, one at a time or in bulk.
No duplicate actions
A repeated or retried action cannot queue or send twice.
Audit log of every action
Records who or what did what, with the result, for the agent and for you.
Per-tool rules in the Sunder agent software
Allow, ask or deny per tool; anything that mixes outside content, private data and a way out always asks. The approval setting the owner picks in the web app is not yet enforced inside the desktop; today it is enforced in the web app and the gateway.
Stop a task, or pause every automatic trigger
A Stop button for each running task and a single switch that pauses all triggers.
Operator can suspend a workspace
SunderSystems staff can suspend a workspace and every request then stops.
Lock settings for a whole team
An administrator locks settings such as sandbox floor, providers and telemetry on every computer that runs our desktop agent. Not offered yet.
Shadow mode before the AI sends
A new AI driver drafts alongside the current one and must match it on at least 90% of 40 messages with no rule breaks.
Staff runbooks for when the AI is down
Forty step-by-step runbooks let people run every job by hand, plus a takeover and a handback.
Client kill switch in the web app
One button stops the AI team and everything it has queued.
Audit export for the client
Downloads the audit log as a file for your own records or auditor.
Privacy and security 9
Private 4B model on our own hardware
The untuned 4B model runs on a GPU we operate, so work on it does not go to a public AI service. It is the fallback when the OpenRouter route is not used.
Choose an open-weight model for a task
Switch a task to an open-weight model through our gateway, with the price shown first. Those requests go to the model's provider.
Training only if you turn it on
Your data is used for training only if an owner of your workspace turns training on in settings, and only for your own model, on our hardware. The choice is recorded and you can turn it off. No per-client model has been delivered yet.
Export your data or ask for deletion
Download your workspace as JSON, or file a deletion request that is tracked to a due date.
Compliance register and incident log
An internal register of controls, a data-request queue and an incident register with a 72-hour clock.
Encrypted nightly backups of the database and of desktop home folders
Backs up the platform database and each hosted desktop's home folder every night and encrypts the copies, which are kept on our own storage. Copies of a deleted desktop age out in about ten weeks. Restoring a desktop is done by us, not from the app.
Mask personal details before a cloud model sees them
Replaces phone numbers and emails before any text goes to a cloud model.
Keys stored in the system keychain
API keys are encrypted with the operating system and never written in plain text.
A dedicated GPU or an on-site box
A model on hardware used by one client only, at our site or yours.
Models and distillation 7
Shared private 4B model with tool calling
A 4-billion-parameter model that reads the inbox and queues actions, at $0.10 in and $0.50 out per million tokens.
Model picker with prices
Pick the 4B or an open-weight model for each workspace, with the cost per million tokens beside each.
Distill a small model for one client's job
Collects approved examples, filters them, trains a small model and switches to it only if it passes a held-out test. Not delivered to any client yet; our first client's own model is in pilot training.
The first per-client model, for HousingInteractive
Moves HousingInteractive's own agent jobs from cloud models onto a model trained on its own data, for its use only, once it passes the safety bar. That model is in pilot training now; delivery is planned and it has not been delivered yet.
A personal adapter for every client
Trains a small adapter on one client's approved work, served from one shared base model.
Hand over your model
Gives a client the weights and the evaluation results to run themselves.
A private AI gateway for your own software
A standard HTTP API in front of your model, with a key and a budget per app.
Access: web, phone, desktop, Telegram 10
Browser app that works on a phone
Chat, approvals, connect, calendar, reports and wallet in a mobile-first web app (invite-only demo).
Approve from your phone
Open the approvals screen in the phone browser and approve, edit or skip each item.
Sunder, the desktop agent
A desktop agent that works in your folders, documents, terminal and browser with approvals.
Run and steer tasks from Telegram
Message your own bot to start, steer and read the result of a task, on paired accounts only.
Drive the desktop agent from other AI apps
Claude, ChatGPT desktop or Cursor can start, follow, steer and stop tasks but cannot approve anything.
Approve the desktop agent from a phone
Follow, steer, stop and approve background tasks through an end-to-end encrypted relay.
Self-serve sign-up
Anyone can start and pay without an invite code.
Installable phone app
Add the app to the phone's home screen, with notifications for approvals.
A hosted desktop for each client
Each client gets its own hosted desktop running Sunder, in a separate container, driven from the web dashboard with a live view and approvals before risky actions. A working prototype that we set up for each client.
Slack and a native mobile app
Run the AI team from Slack and from an app in the app stores.
Nothing matches that combination. Try another area or status.
How control works.
You choose how much the AI may do on its own. In the web app and the gateway the choice is enforced in the software, so a mistake in the model cannot get around it. Enforcing it inside the hosted desktop itself is in progress.
Draft only
The AI drafts everything. Nothing happens until you approve each item.
Ask me first
The AI prepares actions and waits for your yes. Approve several at once.
Act and tell me
The AI does the work and reports back. Every action is logged.
- Approvals you can edit
- Each email, post and booking is a card. Change the wording or the time, then approve. A repeated action cannot go out twice.
- An audit log of every action
- It records what the AI did, what you approved and what happened. It is the same record whether the AI acted or you did.
- Stopping it
- On the desktop agent you can stop any task or pause every automatic trigger. SunderSystems can suspend a workspace. A stop button for the whole team in the web app is planned for Q1 2027.
- Where your data goes
- Your workspace and desktop run on our hardware. AI requests go to an open-weight model: the untuned 4B model on our hardware, or a larger open-weight model through OpenRouter. Requests sent through OpenRouter leave our hardware and go to OpenRouter and the provider it routes to. In the web app, chat and tasks in a new workspace start on a larger open-weight model through OpenRouter. They run on the untuned 4B model on our hardware instead when that route is not available, when the wallet holds no real money or when the plan’s monthly model cap is reached. The AI desktop agent runs on the 4B model on our hardware, or on your own model once one is delivered. A workspace that turns training on moves to its own model once it is trained. That model runs on our hardware.
- Training is opt-in
- Training happens only if an owner of your workspace turns it on. It runs on our own hardware and your data trains only your own model, used for your workspace only, and is never used to train a model for another client. The shared default model is an untuned open-weight model. HousingInteractive, our founder's own company, has agreed for its own data. No per-client model has been delivered yet.
Where it runs.
In a browser, on your phone and on a desktop. Phone approvals work today in the browser. A hosted desktop for each client works as a prototype that we set up by hand. An installable phone app is not built yet.
| Where | Status |
|---|---|
| Browser appChat, approvals, reports and wallet at demo.sundersystems.com (invite only). | Live |
| Phone browserThe same app on your phone: approve, edit or skip each card. | Live |
| TelegramStart, steer and read tasks from your own bot, on paired accounts only. | Live |
| Phone approvals for the desktop agentEnd-to-end encrypted relay, paired by QR code. Needs a relay set up per build. | Pilot |
| Desktop agent for your own computerAn app you would install on your own Windows computer, separate from the hosted Linux desktops. We use it ourselves; it is not offered yet. | Roadmap |
| Installable phone appAdd to the home screen, with approval notifications. | Roadmap, Q1 2027 |
| A hosted desktop for each clientA desktop running Sunder in its own container, with a live view in the web dashboard. Working prototype, set up per client. | Pilot |
Proof: a real operation, with agents drafting the work.
HousingInteractive is a Philippine real estate platform operated by our founder. AI agents draft much of its daily work, with staff approving and covering when the AI is down. Its agents currently use cloud models.
- Checks WhatsApp and Viber every 10 minutes and drafts the replies.
- Invites agents, sweeps about 40 brokers a day and matches buyers who are still searching.
- Sends a daily brief and an availability report, and keeps listings clean.
- Ships website changes through a reviewed deploy, with a backup and an undo for database changes.
- Has 40 written runbooks so people can run every job by hand when the AI is not available.
Those jobs run on larger cloud models today, with phone numbers and emails masked first. A model trained on those jobs, for HousingInteractive only, is in pilot training and has not been delivered yet. HousingInteractive is operated by our founder and has agreed to its own data being used for that. A client's data is never used to train a model for another client, and no per-client model has been delivered yet. The case study has no outcome numbers yet because we have not measured them.
You can also try the live chat on this site, using the button at the bottom right. It answers from this site's own facts on an untuned open-weight 4B model that runs on our hardware.
Integrations.
Names are plain text. We do not use other companies' logos.
| Integration | What it does | Status |
|---|---|---|
| Email (sandbox mailbox) | Read customer emails and send replies in the demo workspace. | Live |
| Amazon SES | Send mail from your domain with bounce handling. | Pilot |
| Resend | Send website form and system mail. | Pilot |
| Amazon WorkMail | Read the shared team mailbox in the portal. | Roadmap, Q1 2027 |
| SendPulse | Send campaign email to a list. | Roadmap, Q2 2027 |
| Check and answer chats through a browser tab. | Pilot | |
| Viber | Check and answer chats through the desktop app. | Pilot |
| Telegram | Run and steer desktop-agent tasks from your own bot. | Live |
| Messenger | Read and answer page messages. | Roadmap, Q2 2027 |
| Facebook Page | Publish approved posts to a sandbox feed. | Live |
| Publish approved posts to a sandbox feed. | Live | |
| Publish to a sandbox feed; company pages need partner approval. | Live | |
| Google Calendar | Book meetings on your own calendar. | Roadmap, Q1 2027 |
| Google Analytics | Read traffic and channel figures. | Pilot |
| Supabase | Read tables without being able to change them. | Live |
| Amazon Web Services | Read EC2 and related resources; deploys run per client. | Pilot |
| GitHub and GitLab | Review diffs, open pull requests, fix failing checks. | Live |
| Jira | Read and answer tickets. | Pilot |
| Stripe | Wallet top-ups and subscription, test mode today. | Live |
| MCP servers | Add any Model Context Protocol connector to the desktop agent. | Live |
| Twilio or Telnyx | Phone numbers for the voice receptionist. | Roadmap, Q3 2027 |
| Meta Ads and Google Ads | Draft ad copy and manage budgets. | Roadmap, Q3 2027 |
Ten founding places. Your price is locked for 12 months.
Founding clients get done-for-you setup of the Live and Pilot items they need, and a direct line to the founder. In return we ask for a case study after 60 days.