The model stays local. Other routes may not.
Standard agent permissions — permitted tools or scripts may still connect.
Hire the model. Keep control of the work.
Agenaxy is a local-first AI agent workbench for automating real work with documents, datasets, and recurring workflows.
Some work should never end up on someone else’s server — that’s what Vault Mode is for. It works only with models you trust. An OS-level sandbox restricts the agent to approved files and folders and prevents data from being sent elsewhere, keeping every run within the boundaries you define.
Source files, Activity, and finished Artifacts stay visible in one workbench. Switch between the separate Standard and Vault chat modes to inspect both.
Who this is for
You probably need Agenaxy when the file is why you’re still doing it by hand.
Five moments you may recognize before the first run.
- 01
You deleted your customers’ names and addresses before you dared hand the file to a chat.
Hand over the whole file. The model runs on this Mac, or on a server you trust — the list never reaches an outside provider, and can’t be passed on from there.
- 02
The quarter isn’t announced yet, and the reconciliation is sitting in four spreadsheets.
Run it in Vault with a model on this Mac and the figures never leave. Nothing is sitting on someone else’s server when the announcement goes out.
- 03
You’ve signed something that says “don’t paste this into a third party.”
Use a model the paperwork already covers — one on this Mac, or a server you trust: your own deployment, or a partner the contract already names. Vault allows that one connection and closes every other route.
- 04
The draft isn’t published and the interview isn’t cleared — and you’d rather neither became training data.
Unpublished work stays a local file. Nothing is uploaded, so there is no retention policy to read and nothing to opt out of.
- 05
You don’t want to give an agent write access to a folder you couldn’t rebuild.
Every run works on a copy of your files, so the originals are never what’s at risk.
The platform way
Every device is a door into a hosted workspace. What keeps your work private there is a promise — and promises fail in ordinary ways: a bug that hands your chats to a stranger, a subprocessor you never heard of getting breached, terms that change after you agreed to them, one line of misconfiguration that puts your conversations in Google.
The Agenaxy way
The workspace is yours; the model is hired into it, task by task. Your files, chats and results never enter a hosted workspace, so none of those failures has a route to them. What holds here isn’t a policy someone has to maintain — it’s a boundary the operating system enforces.
Complete runs · different kinds of work
Hand off real work. Get an artifact you can defend.
Confidential, regulated, personal, or unpublished — the work that still gets done by hand because it can’t be sent anywhere. Each goes in with a plain-language task; a finished, source-linked artifact comes out, on your Mac.
E-commerce ops selected
Source · on this MacCustomer PII
orders.csvrefunds.csvreviews.csvsupport.csv
Task · plain language
“Turn last week’s exports into a Monday brief. Reconcile the numbers, flag what needs a human — don’t guess causes.”
Run · local model
- 1Join & reconcile rowscode
- 2Group & draft signalsmodel
- 3Flag what needs a humanreview
Artifact
weekly-ops-brief.mdsaved locally
Fulfilled orders612↑ 7.6%
Gross sales$52.4k↑ 5.2%
Refund rate2.9%↑ 0.8 pp
Support84↑ 16.7%
Fulfilled orders by day
612 current · 569 priorTop support topics
38 taggedCalculatedRefund rate ↑ 0.8pp
18 ÷ 612 = 2.94%
Model-draftedEast delivery cluster
9 of 14 late-delivery from East. Cause not established.
Human-approvedInspect 12 orders
Carrier scans by Tue · Ops
What stays · what leaves
Choose the model. Set the boundary.
Agenaxy runs in two modes. Choose a model on this Mac or on a server you trust. Standard keeps normal agent permissions; Vault allows only the model connection you chose, and stops the agent from sending your data anywhere else.
Task context goes to the model provider.
Standard agent permissions — task context goes to the provider; tools or scripts may still connect.
Nothing leaves this Mac.
Vault rules enforced by the system sandbox — data-sending tools off, script network blocked.
Only selected context reaches the server you approved.
Vault rules enforced by the system sandbox — every route except the server you approved is blocked.
Local Workbench · system of record
- Everything stays here. Source files, artifacts, chats, and the full run trace remain stored on this Mac in every configuration.
- Your originals stay untouched. Every run works on a copy. Whatever the agent does, the files you started with are exactly as they were.
- You choose what leaves. Pick a remote model for a task and it gets that task’s context — never your history, never the rest of this Mac.
Run record
One run, refusals included.
The request, every step, the connection that was refused, and the finished file all stay in one place — on this Mac.
The scanned addendum needed an OCR service outside this Mac — the one page you’d least want to upload. The sandbox refused the connection, so it stayed here, and the run finished with the two text agreements.
Confidential agreement review
VAULT ON
TaskReview these confidential agreements — two text documents and one scanned page. Flag risky clauses and produce a comparison table.
- Read the local files Running Done
- Local model analyzes the clauses Running Done
- OCR tool asks to send the scanned page Requested Blockedocr-extract→third-party-ocr.comSandbox: ocr-extract(41287) deny(1) network-outbound api.ocr-cloud.io:443
connect: Operation not permitted (errno 1) - Comparison table built from the text agreements Running Done
contract-comparison.md
Two agreements compared · the signed addendum stayed here
Agenaxy for Mac
What would you put in the Vault?
Tell us about the file or workflow you need to keep under your control.
Try AgenaxyGuides for the work behind the work
Learn the boundary. See the workflow.
Start with the concepts behind local-first agents, or jump to a complete use case for documents, finance, legal work, ecommerce, content, and small teams.
Learn
AI agent basics, private AI, Mac models, document workflows, reliability, and comparisons.
Browse learning guides →Use cases
Concrete examples of local-first AI producing work that remains reviewable and under your control.
Browse use cases →