LARS · Local Agentic RAG System
Your AI. Your hardware. Your data.
LARS is a private AI system we build around your company and run on hardware you own. It reads your documents, answers with citations, writes the report, runs your calendar and inbox, and never sends a word to an AI provider.
The real LARS interface, shown with sample documents.
LARS by the numbers
0
Prompts sent to an AI provider
Chat, embeddings, vision and transcription all run on the box.
8
Conversations at once
On one 128 GB box, sharing a single model load.
512K
Tokens of working context
Up to 65K per conversation.
6
Local AI models
Chat, embeddings, reranking, vision, titles and speech-to-text.
1.8M
Spreadsheet rows indexed
In a production library, searchable row by row.
A better way to do AI
Every prompt you send to the big AI companies leaves the building and comes back as a bill by the token. LARS flips both. Your data stays on a box you own, and a question costs the same whether your team asks ten or ten thousand.
One assistant for the whole office.
Two agents share the work: LARS for fast, grounded answers from your files, and Hermes for longer multi-step tasks. Both run on the same box.
Reads
Every file you have
PDF, Word, Excel and PowerPoint (old formats too), email, HTML, EPUB, scanned images through OCR and vision, Photoshop files, and audio or video through transcription. Every file gets a summary, not just an index entry.
Answers
With the source attached
Search across your library with reranking, read spreadsheets row by row, and do the math with an exact calculator instead of guessing. Every answer cites the file it came from.
Builds
The deliverable, not a draft
Documents out to PDF and Word, spreadsheets out to Excel, and interactive dashboards that render in a locked-down sandbox. Precise edits keep the rest of a document intact.
Schedules
Your calendar
Day, week, month and year views with an assistant built in. It checks your schedule, finds free time and books meetings, and connects to iCloud, Fastmail or Nextcloud calendars.
Answers mail
Your inbox
Connect IMAP and SMTP accounts. LARS reads, searches, triages and sends, with attachments pulled from your files or its own reports. Passwords are encrypted and never shown to the model.
Automates
Agent Flows
Multi-step workflows that run on a schedule and produce cited reports: the Monday numbers, the daily digest, the month-end summary, delivered by email before you sit down.
Talks
Team messaging
Direct messages, groups and channels with @LARS in the conversation, under a strict rule about what it may read from them.
Extends
Tools your team builds
Ask in chat for a reconciliation, a variance check or a report generator. Tools run in isolated sandboxes with no network, can require approval before they apply, and become callable by LARS.
Shows
Your knowledge as a map
A 3-D map of every document by meaning. See what clusters together, find near-duplicates and pull a set of documents straight into a chat.
Run the numbers
Stop paying by the token.
Agent work is token-hungry: a multi-step answer on LARS reads 45K to 106K tokens of documents and tool results. On a frontier API, every one of those is metered. On LARS, they’re free once the box is in.
Requests a month
7,875
480.4M tokens
Claude Opus 5.5 API
$1,525/mo
$18,305 a year, billed by the token
LARS Studio
$111/mo
Hardware over 3 years + ~$11 of power
Pays for itself in
2.4 months
Then you keep $16,973 a year
Three years, side by side
Compute only: setup, customization and support are quoted per company. API prices are Anthropic list prices as of October 2026, with 70% of input billed at the cached rate to keep the comparison fair. 21 workdays a month, power at $0.15/kWh. Not counted: document summaries, embeddings, OCR, transcription and scheduled flows, which LARS also runs locally at no per-token cost.
Sizing guide
Sized to your team.
Same software at every size. Start on one box and move up without starting over. We confirm the fit on your own workload before you buy.
01
In production todayStudio
Up to ~40 people
- Hardware
- AMD Ryzen AI Max+ 395 workstation, 128 GB unified memory
- At once
- 8 conversations at once
- Models
- 35B mixture-of-experts chat model, plus embeddings, reranking, vision and speech
- Hardware cost
- $3,300 to $3,900
02
Sizing estimateOffice
~40 to 150 people
- Hardware
- Workstation with an NVIDIA RTX PRO 6000 (96 GB)
- At once
- 16 to 32 conversations at once
- Models
- About 7× the memory bandwidth for faster answers, with room for larger models
- Hardware cost
- $13,000 to $17,000
03
Sizing estimateEnterprise
150+ people
- Hardware
- Multi-GPU server, two to eight GPUs
- At once
- 64+ conversations at once
- Models
- The largest open models, several running side by side
- Hardware cost
- Quoted
Team sizes assume about one in five people has a request running at peak. Every tier stores documents on its own disks; add storage, not seats.
Private by design
Nothing leaves the building.
Built for firms that can’t paste client files into a chatbot: financial, legal, medical and anyone under an NDA.
Offline by default
Chat, embeddings, vision and transcription run on local models. There is no frontier provider in the loop.
Web only when you say so
Public web lookups need the admin switch on and your Allow in the chat, and can only reach public addresses.
One switch cuts it off
An admin kill switch stops all outbound connector traffic instantly.
Encrypted credentials
Email and calendar passwords are encrypted at rest and never returned to the browser or the model.
Roles, approvals, audit
Admin, Builder and Member roles, approvals before actions apply, and an audit log of who did what.
Backed up nightly
People, chats, messages and settings are backed up every night, with alerts if anything fails.
Not off the shelf
Built around how you work.
01
Your name on it
Each deployment carries its own name and branding. Your team uses your assistant, not ours.
02
Your company context
A per-project memory of what you do, how your folders are organized and what each document is, so answers start oriented.
03
Your playbooks
Skills teach LARS how your firm does things: how you analyze a statement, write a variance memo or answer a client.
04
Better every week
Every answer is logged and scored, problems surface in a weekly training book, and fixes are proven on a regression suite before they ship.
01
Size it
We look at what your team asks every day, how many people ask at once and how much you store, then pick the box.
02
Install it
The box goes in your office or rack. We connect your file share, email and calendars, and load your documents.
03
Shape it
Your name and branding, your company context, your playbooks and the flows you want every morning.
04
Run it
It updates itself, backs itself up and tells us when something needs a look. You call one number.
FAQ
Questions, answered.
Yes. The chat model, embeddings, reranking, vision and transcription all run on the box. Nothing is sent to OpenAI, Google, Anthropic or anyone else to produce an answer. Public web lookups exist, but only with the admin switch on and your Allow in the chat, and an admin can cut all outbound traffic with one switch.
The biggest frontier models are still stronger at the hardest open-ended reasoning. LARS is built for company work: your documents, spreadsheets, inbox and calendar, with citations and exact math. Every change is measured on a regression suite of real questions before it ships, so it gets better at your work week to week.
Most teams up to about 40 people start on a single 128 GB AMD workstation, the same box LARS runs on in production. Larger teams move to a GPU workstation or a multi-GPU server. The software is the same at every size, so you can start small and move up without starting over. We size it on your workload before you buy.
You own the hardware, and there is no per-token metering: a question costs the same whether your team asks ten or ten thousand. Power for the Studio box runs about $10 to $15 a month. Setup, customization and support from Cinderwire are quoted for your company.
Drop files into a shared network folder and LARS ingests them automatically, or upload them in the app. Two-way sync with Dropbox, OneDrive, SharePoint and Google Drive is built and in testing. Personal files stay private to each person.
We do. The box updates itself within minutes when we ship an improvement, backs itself up every night, and alerts us if an update, a backup or the AI model fails. You get one team and one number to call, the same as the rest of your IT.
LARS demo
Bring your AI in-house.
Thirty minutes: we look at what your team asks AI every day, size the box, and show you LARS working on real documents.