Connections are quoted, not assumed
Each supported integration, permission boundary and acceptance test is written into the quote. It is not treated as delivered until it passes on the final machine.
The LocallyAI Harness
Your whole working day,
in one browser tab.
The Harness is the software half of what we sell. It runs on the machine we install in your office and brings your documents, meetings, research, tools and standing jobs into one place — with the everyday model running on your own hardware.
- Installed on hardware you own
- No per-user Harness fee
- Base offline work stays on your appliance
Watch it research a question, then chart the answer — one thread, nothing trimmed
Rotate your phone for the full desktop view.
Start here
The six features to test first.
The complete list is below. Start with these six because they cover common office work and are straightforward to test against agreed examples before you buy.
-
01 · Documents and numbers
Answers with the working shown
Point it at the folder your spreadsheets already live in. The answer comes back with the steps, the source file and a chart.
-
02 · Deep research
The question broken into its parts
One question becomes seven lines of enquiry, each showing how many sources answered it, ending in a cited report you can send.
-
03 · Meetings
Decisions and actions, with owners
Record on a PC or a phone on the office wifi. Transcription runs locally, and the meeting becomes a searchable source of its own.
-
04 · Project handovers
Records turned into a finished pack
Bring drawings, meeting notes and defect records together. Open items stay visible and every conclusion links to its supporting record.
-
05 · Scheduled work
Standing jobs that just run
A morning brief, a Monday digest, an aged-debt list. Set once in plain English, running on the appliance whether anyone is signed in or not.
-
06 · Cited document answers
The source stays beside the answer
Ask against approved project files, then inspect the clause, variation or record behind the response without leaving the conversation.
Everything else — and there is a great deal of it — is listed below, section by section.
Working core
Ask about a contract.
Ask about a contract.
Check the source.
The dependable path today is chat grounded in office material, with the source kept close enough for a person to review.
Chat
Streaming replies, conversation history and search, attachments, slash commands, an agent picker and three staff presets all live in the main conversation surface. The Harness loads the relevant skills for the job and knows the current date.
Image-capable models can receive image attachments. A text-only model refuses cleanly instead of pretending it saw the image. Reasoning output stays separate from the final answer when the model provides it.
Document answers
Upload a file, let the Harness index it, then ask a question. Your own documents rank ahead of general material and the answer links back to its source.
PDF, Word, Excel, PowerPoint, CSV, text and Markdown are all read and indexed.
The file room carries semantic and exact search, a source reader, index status and watched folders. Files can be pasted in, filed, shared, revised or deleted, and every change is recorded. A file that fails to read says why, rather than turning up empty.
Workspaces
Keep the work with the matter.
The aim is one place for the question, source material, notes, outputs and people allowed to see them.
Matters and folders
A matter picker and tree organise work by client, job or file. Groups, ordering, drag and drop, menus and dialogs support ordinary folder administration. Work can be personal or shared without making every private conversation public.
Notebooks
A notebook scopes a set of uploads, pasted text, meetings and notes. It can be private or shared, allow one writer with live readers, turn notes into sources, suggest follow-up questions and create studio outputs or a Word export.
Files
Search and clear search, change tabs, drag or upload files, watch the queue, open the reader, share, delete and manage folders without leaving the central workspace.
Meetings
Record, transcribe, then discuss.
Record in the browser or on a phone on the office wifi. Keep the tab open while it runs — closing it ends the recording.
Meeting controls cover record and upload from a computer or phone, consent, pause, resume and discard. Private and shared history and detail views hold the recording, transcript, summary and actions, with live status, search, chat, sharing, export, retention and deletion around them. Dictation uses the same local speech service.
In the standard local transcription configuration, recordings from a browser or phone are converted and transcribed on the appliance with Whisper large v3 turbo. A connected capture or storage workflow is treated as a separate exception and named in the quote.
Deep research
Show the plan before the long run.
Research can work across permitted documents, notebooks and meetings, with optional web access through the local LocallySearch service.
Start from chat or a notebook. Refine the question, inspect or edit the plan, choose whether the web is allowed, approve or discard it, then follow workers and sources inline. You can stop a run, read the report, check citations, export it to Word and share or delete the result.
Deep research plans the question into parallel lines of enquiry, reads across your library and the web where allowed, and writes a cited report you can send.
Tools and integrations
Keep repeat work close to the chat.
Small tools
Calculators, formatters and checkers can be fitted to the office library. The written scope names what staff may run, where output is saved and how it is checked before handover.
Skills and office integrations
Skills guide document, spreadsheet, presentation and PDF work. Industry packs add the written rules and starting material for a field. MCP connections and a managed integration pane cover systems an office chooses to connect.
Mailbox access is a fitted integration, not a blanket base feature. If it is included in the quote, each mailbox, permitted action and approval step is connected and proven during the install’s acceptance checks.
Mini environments and CLI
Where a scoped workflow needs Python, we fit a restricted workspace and prove the permitted inputs, outputs and file access before handover.
The LocallyAI CLI is available for developer workflows on macOS, Windows and Linux where it is included in the written scope and acceptance test.
Scheduled work
Say it once. Let it run.
A standing job is asked for the same way as anything else — in a sentence, in the chat. The Harness works out what it needs to reach, creates the schedule and shows you the first run before you commit to it.
Jobs, cadences and the run list
Each job carries its own cadence — weekdays at 09:00, every week at 07:00, twice a day, monthly — along with what it does, who owns it, when it last ran and when it runs next. Jobs run on the appliance whether or not anyone is signed in.
Every job can be run on the spot, edited, switched off or deleted. A job that writes to disk carries a warning on its card before you enable it, rather than after.
Typical jobs: a per-person morning brief, a Monday sales digest against target, an overnight summary of unread mail, a list of invoices past sixty days with who chased them last, and monthly filing or archive runs.
Accounts and administration
Useful controls, without reading everybody's work.
Accounts, roles and separation
The account scope covers first-boot setup, local sign-in, sessions, forced password changes, people administration and per-user capabilities. Shared office areas and directory allowlists are deliberate configuration choices.
For a multi-user or sensitive deployment, the quote names the required separation controls and the acceptance test proves them before staff data is loaded. A control that has not passed that test is not represented as active.
Admin record and health
Administrators manage people, network blocks and the operational record. Their access boundary is recorded and tested for the deployment rather than assumed from a role label.
The admin view also covers usage, health checks, scheduled briefings and benchmark information configured for that install.
Records, export and models
Keep a trail. Keep a way out.
File notes and exports
The system keeps a trace of source, instruction and output, with administrator control over what is retained. File notes, exports and research reports can be reviewed from that same record, and the agreed export path is tested before handover.
Training data and fine-tuning
Approved work can be exported as a training set. Fine-tuning may run on site or, if separately requested, off site. Off-site fine-tuning is a named data-leaving exception. Rights in your material and the licence for any base or third-party model remain separate matters.
Model handling
Models load when needed, release after three idle hours and use prompt caching, so the machine is not burning memory on work nobody asked for. Warming keeps the everyday model ready, and fallback covers the case where a larger model is busy. Your quote names the context length the chosen tier carries.
Installation and support
One office system, fitted and recorded.
Local access
The appliance serves the Harness over HTTPS on the office network at a local address. Staff use a browser on a computer or phone; the ordinary model, transcription, documents and saved tools stay on the appliance in the base setup. Staff accounts are not licensed one by one.
The interface uses plain Australian English, with a simpler surface for staff and a fuller technical surface for administrators. It works on a phone on the office network for recording and dictation.
Updates and support
Improvements arrive through the support plan rather than a new licence. Applying a Harness update needs a temporary connection for the update and checks, then that connection is closed again.
Web research, updates, external integrations, remote support and off-site fine-tuning are all separate exceptions to base offline operation. Each chosen exception belongs in the quote and configuration record.
Own the working system
Why businesses choose LocallyAI over ChatGPT or Claude.
The Harness puts approved routine work, local documents, tools and generated files inside one browser workspace running against hardware your business owns.
Quoted automations run on the appliance
A morning brief, weekly digest or aged-debt list is included only when named in the quote and proven against agreed examples before handover.
Processing location is the difference
Hosted services require your business to rely on a provider’s infrastructure, terms, configuration and continuing service. LocallyAI keeps approved base workflows on your appliance; connected exceptions are documented and acceptance-tested.
A boundary your team can inspect
The standard workflow is local. Web research, integrations, remote support, updates and off-site backups are separate exceptions recorded in the quote and tested before handover.
Core work can continue offline
Local chat, approved document search and transcription can keep running when the connection drops. Web research and connected services cannot.
Built from scratch, for this
We did not rebrand a coding interface. The Harness is written for office workflows and fitted to the templates, letterhead and approved systems named in your quote.
The Harness is modular
Supported skills and MCP connections can be added without replacing the interface. Each requested module still needs a written scope and delivery test.
You can actually talk to us
We build the software, so a requested change can be discussed with the people who write it. Standard product improvements and business-specific work are separated in the written scope, including any cost.
Why a Harness at all
A model is a horse. The Harness is the saddle.
The models everyone talks about are horses — powerful, and useless to a business on their own. The saddle is what makes one rideable: the software that straps it to your documents, your templates, your rules and your people. Nearly every saddle ever built was made for coding. This one is built for running a business — and everything on this page is what that means in practice.
Whole conversations, start to finish
What using it actually looks like.
Three complete threads, nothing trimmed out of the middle. Each one starts with a question somebody would really ask and ends with a thing you can send, file or switch on. Scroll inside a frame to read the whole exchange.
Example 01
“Can you build me a quick quote calculator?”
A builder wants pricing that stands up in New South Wales. The Harness checks two industry knowledge packs before it writes a line of the tool, then produces the quote itself.
- The question, in the words a builder would use: materials at cost plus a margin, labour at $95 an hour.
- It reads two industry knowledge packs — 3,800 sources between them — before building anything.
- The tool appears in the conversation with live fields and a total of $15,710.20 including GST.
- It names the four rules it had to honour: GST at 10%, the licence number, the written-contract threshold and the 10% deposit cap.
- It offers to save the tool to the company library, then writes the finished quote as a PDF from the office template.
Expand and scroll to see the below transcript
Example 02
“Which suppliers put prices up more than 5%?”
A finance lead has a board meeting on Thursday. Deep research plans the question into seven lines of enquiry, sweeps the company file store and the web, and writes a cited report.
- One question in, seven separate lines of enquiry out — each one shown with the number of sources behind it.
- It reads 84 sources: 74 from the finance share, the nightly ledger export and company policies, 10 from the web.
- It reconciles 1,204 purchase lines against nine suppliers’ agreed rates.
- The answer leads with the number for Thursday: $186,000 added to annualised cost of sales, $54,000 of it recoverable.
- The full report is written to the chat as a file, with every figure cited back to the invoice or clause.
Expand and scroll to see the below transcript
Example 03
“Every weekday at 9, I want a morning brief.”
Asking for a standing job is the same as asking for anything else. The Harness checks what it can actually reach, creates the schedule, then shows what this morning’s brief would have said.
- The request in plain English, including what counts as worth mentioning.
- It checks what it can reach first — mail, the company file store, the task list — rather than assuming.
- The schedule is created for weekdays at 09:00, sandboxed, with no disk writes.
- It runs the brief immediately against today’s real data so you can see the shape of it before tomorrow.
- The job card appears in the thread with a toggle, and again under Scheduled in the sidebar.
Expand and scroll to see the below transcript
Shared improvements are not billed as bespoke work.
If a requested improvement belongs in the standard Harness, the quote can treat it as a platform change rather than customer-specific development. Business-specific integrations, rules and deliverables are scoped and priced separately.
These are the demonstration conversations we walk through on a call, run against example company data rather than yours. Book a walkthrough and we will point it at a document of yours instead.
Bring us one job your team repeats every week. We will show you the system doing it →