Secure AI for local businessSecure AI for local and blue-collar companies
In the cloud on the frontier models, or on a machine in your building that never phones home.
Lost in the weeds with AI, or dreaming of your own and not sure where to start? Maru builds it: agents wired into the tools you already own, running on GPT, Claude, Grok and Gemini, or on a private box in your back office with nothing leaving the building. Ten business days, fixed price, written security audit.
Agents wired into the tools you already own, on the frontier models or on a private box in your building. Ten business days, fixed price, written security audit.
01/1414 nodes · whole workflow
p50 latency
-
tool calls
0
tokens
0
cost
-
handoff
watching
Two ways to run it
Every build runs one of two ways, and the agents are identical either way. The only decision is where the model lives and where your data is allowed to go. We will tell you which one you actually need on the first call.
In the cloud
GPT, Claude, Grok and Gemini through one gateway you own. You pay the providers directly. No markup, no middleman account, no lock-in to whichever lab is winning this quarter.
For companies that want the strongest models and are fine with provider terms.
Included in every build
In your building
A Mac mini or a Ryzen AI Max box on your own network, running an open-weight model in the 27B to 70B class. The same agents, the same tools, and not one byte of client data leaves the building. For law, medical, the trades, and anyone who has been burned once.
Quick Win pricing. Hardware is yours outright.
From $6,500 installed · the kits →Or both
Most companies end up hybrid. Client files and anything regulated stay on the box in the back office; everything else uses the frontier. One system, one audit log, and the agents cannot tell the difference.
Runs inside the tools you already pay for
The actual problem
That is the layer the agent belongs in. Not a chatbot on the website. Software that does the job the copying-and-pasting person is doing right now, inside the tools you already pay for.
ChatGPT wrote a nice email. It cannot see your calendar, your invoices or your customer list, so the person copying and pasting is still copying and pasting.
A missed call at 6:40pm is a booked appointment somewhere else by 6:45. The voicemail gets checked in the morning, if it gets checked.
The scheduler doesn't know what the invoicing does. So a careful, expensive person spends Thursday copying numbers between two screens. They are the only reason the figures line up.
Nobody wants to touch it. The person who built it may not work there anymore. It has no backup and no owner, and it decides who gets paid.
It doesn't. A 27B model on a box in your back office does most of this work now, and nothing has to leave the building.
Three of the five and the build is worth a conversation. All five and it is worth one this week.
Anatomy of an agent
Not a chatbot with a phone number. A trigger, a model held to a schema, memory in your database, scoped tools, guardrails that block before a write, and a person at the end of the line. Every layer writes to the same log.
Pricing
Fixed price, fixed scope, on paper before anything starts. What is in and what is out are both listed. You sign one page, not a statement of work with a change-order clause.
One thing, automated
$3,500 to $6,500
In production in ten business days
Pick the one thing that eats your week and we take it off your desk. One agent from the catalog, wired into the tools you run, tested against your real data.
Pick one
Ships with. A written security note: what the AI can see and where that data lives.
When one specific thing is eating a person's week.
Talk it throughThe system it all runs on
$15,000 to $25,000
Three to six weeks, scoped in writing
Website, client database, and scheduling, ordering, repairs or invoicing, with the Quick Win agents wired through all of it. For companies whose current stack is a spreadsheet and three tools that disagree.
Everything in one place
Ships with. A written security audit against our 136-control methodology.
When the answer to 'where does that live' is four different places.
Book a fit callWith every build
$600 to $1,500 a month
Starts on handover day
Agents break when a vendor changes an API. This is the plan that catches it before you notice, which is why it comes with every build rather than being optional.
Included
Ships with. Cancel with thirty days' notice. Your instance and data stay yours.
Required with every build. Priced by what is being watched.
Talk it throughAnything not on the catalog is quoted as a custom build with its own honest timeline. The offline box is Quick Win pricing with the hardware included and yours outright. Multi-location, multi-language and anything that takes payment over the phone are quoted separately.
The catalog
An agent joins the shelf once it runs in production. The one that booked this call is running on this page. Each listing says plainly whether it is running, ready in ten business days, or a first build we scope individually.
Answers your phone, qualifies the caller, books the appointment, logs it.
Twilio Voice · calendar · CRM
$6,500
A form that qualifies the lead and books it, instead of just collecting it.
web · SMS · calendar · CRM
$3,500
Keeps two systems that don't talk in agreement, continuously.
POS · accounting · SQL
$6,500
Follows up on invoices and estimates until they're paid or declined.
QuickBooks · Xero · Stripe
$4,500
The numbers you can't see until the 15th, in your inbox every Monday.
any source with an API
$3,500
A missed call becomes a text conversation in under sixty seconds.
Twilio Messaging · calendar
$3,500
Each price is the Quick Win price for that agent. Care from $600 a month.
How ten days actually works
It starts when credentials are delivered and the spec is signed, not when you first email. That is the only honest way to promise a date. It is possible because proven components get assembled rather than invented. The ten-day clock applies to agents marked ready; a first build of Front Desk or Reconcile is scoped on its own timeline, in writing.
Before the clock
Thirty minutes. We establish whether this is worth building. You leave with the two things I would automate first, whether or not you hire me.
Before the clock
A written spec with a fixed price. In scope and out of scope are both listed. You sign one page.
Day 0
The clock starts when credentials are delivered and the spec is signed. Not before.
Days 1 to 7
You see it running mid-week, not at the end. Nothing is a surprise on delivery day.
Days 8 and 9
Not a demo dataset. Your records, your edge cases, your worst-behaved customer.
Day 10
Documentation, the security note, and the keys. It runs in your accounts. You can revoke my access in one click.
Security posture
Almost nobody selling AI to local businesses will say that out loud. The real question about handing software the keys is not whether it works. It is what happens to you if it doesn't. That gets answered in the architecture, before it gets answered in a contract.
Your accounts, not ours
Your automation platform, your database, your model API key, your phone number. I build it, configure it, and hand you the keys.
We don't hold your data
Nothing is stored on our infrastructure. There is no copy of your customer list sitting somewhere waiting to be breached.
Revoke in one click
Access is delegated and reversible. The day you don't want me in there, you're not in a support ticket, you're in a settings page.
Every build ships with an audit
A written report on what the agent can reach, what it can change, and where that data lives. It is the document your insurer will eventually ask for.
Sample report · 21 pages · redacted engagement · ungated
What we've built
Below is the Business OS reference architecture: intake and phone, money, repairs, operations, and the role-based dashboards behind all of it. It is drawn from a system running in production for a client, with the identifying detail removed; every node is a component we build from. Then the layers underneath: security, memory, and the guard that keeps agents inside their lane.
Website, client database, repairs, consignment and invoicing for a fine jeweler, replacing a spreadsheet and three tools that disagreed. In daily use by the owner and staff.
A production application holding to twenty-seven of twenty-seven adversarial checks, and a 136-control audit methodology run against our own builds before anyone else's.
Cross-model memory infrastructure and MCP servers: persistent context that survives between sessions and between vendors. The plumbing agents need to remember anything at all.
Six AI assistants on one shared memory store, sixty-three skills, and a pre-execution hook that hard-blocks dangerous tool calls before they run. The kill switch, built before it was needed.
Who builds it
Trained in
Software and security. Certified, and proven in production, not a bootcamp pivot.
Ran
A multi-location operating business. P&L, staff, corporate reporting.
Builds
The software the business runs on, and the audits that test it.
Based
Remote across the United States, on Eastern time.
Most people selling you AI have never run a business. Most people who have run one can't build. That combination is the whole offer.
I trained in software and security and then went and ran a business instead. Years of it: the P&L, a corporate parent to answer to, and the software nobody in the building liked.
Reconciling systems that disagreed. Watching a good person spend every Thursday moving numbers between two screens. Eventually I stopped complaining about it and built our own. Then people started asking me to build theirs.
That is why the first call is worth thirty minutes even if you never hire me. I will tell you what I would automate first, and I will tell you if the answer is nothing.
Questions
It's the point. Agents that live outside your stack are chatbots. The value is in the plumbing: reading from the scheduler you already pay for, writing to the accounting system you already reconcile, booking into the calendar your staff already watch. If you had nothing, this would be a much smaller job and a much smaller return.
Yes. A Mac mini or a Ryzen AI Max box on your network runs an open-weight model in the 27B to 70B class, and the same agents run against it. Nothing leaves the building and it keeps working when the internet doesn't. From $6,500 installed with the hardware included and yours outright. Most companies end up hybrid: regulated work on the box, everything else on the frontier models.
Whichever is best for the job, on your keys. GPT, Claude, Grok and Gemini in the cloud, routed through a gateway you own with failover and spend caps; or Qwen, Llama and Gemma on a box in your building. The agents do not care which model is behind them, which is the point. When a better one ships, it gets swapped in under the care plan.
For a single agent on the catalog, yes, and the clock is defined: it starts when credentials are delivered and the spec is signed, not when you first email. It is possible because proven components get assembled rather than invented, which is also why anything not on the shelf gets quoted with its own honest timeline.
You do. It runs in your accounts, on your keys, against your data, and the documentation is written for someone who is not me. You keep your instance and your data outright. We retain the right to reuse generic components and methods, which is standard and is what keeps the catalog moving.
No. Your automation platform, your database, your model API key, your phone number. Nothing is stored on our infrastructure, so there is no copy of your customer list sitting somewhere waiting to be breached. Access is delegated and you can revoke it in one click.
It will, eventually, because vendors change APIs and models change behaviour. That is why the care plan comes with every build rather than being optional. Monitoring catches the failure before you notice it, and a fixed block of change hours covers the fix. An unmonitored agent that silently stops working in month six is worse than no agent at all.
No. You need to know your business well enough to say what eats your week and who does it by hand. The technical detail is published on every listing for the people who want it, and safely ignored by everyone else.
Yes, and it usually goes well, because this is the work they typically do not have time for. Scope is written down before anything starts specifically so there is no ambiguity about who owns what.
Because an agent with credentials to your systems is a new attack surface, and almost nobody selling them will tell you that. Every build ships with a written report covering what the agent can reach, what it can change, and where that data lives. It costs you nothing extra and it is the document your insurer or your biggest customer will eventually ask for.
Thursday and Friday afternoons · weekday evenings · ET
Thirty minutes. I will have looked at how your operation runs before we talk, and you will leave with the two things I would automate first, whether or not you hire me.
The form that books this call is Intake, from the catalog. The product is running on this page.