Local-first AI agent · macOS · Windows · Linux
An AI agent that does the work on your own machine.
Halo organizes files, drafts documents, cleans data and runs the chores you would rather not do. Every request starts on a private local model — free, offline, no API bill. The cloud is a fallback you approve, never a default.
$0
per local request
0
files uploaded
7
installers, all platforms
01 · Local first, by design
Every request starts local.
Halo gives your local model the job first and only reaches for the internet when the task genuinely needs it — and never without asking. For most people that means most days cost nothing and need no connection at all.
01
Start local
The request goes to the model already installed on your machine. No key, no meter, no upload.
02
Stop if it cannot
If the local model cannot finish with confidence, Halo reports what it found and does not spend anything.
03
Escalate on your word
You decide to send that one request to a cloud model, on your own key and inside your own spending limit.
| Local model | Cloud model | |
|---|---|---|
| Cost per request | $0 | Billed per token |
| Internet | Not required | Required |
| Your files | Never leave the machine | Uploaded to a provider |
| Setup | Ollama, one command | API key and billing account |
| Best for | The everyday ninety percent | Deep reasoning, live web data |
Local model means a model such as Qwen, Llama or DeepSeek running on your own computer through Ollama. Cloud model means a hosted provider billed per token, on a key you supply.
02 · Ask it locally
Twenty-four things you can hand Halo today — no key, no bill.
Each example below runs on a local model with the network switched off. Say them the way you would say them to a colleague.
Files and folders
Housekeeping that never needs a server.
“Sort my Downloads into folders by type and date.”
Reads every file, moves it, and shows you the plan first.
“Find the duplicate photos and keep the sharpest copy.”
Compares by content, not filename, then reports the space freed.
“Rename every invoice to Invoice-YYYY-MM-DD-Vendor.”
Opens each PDF to read the date and the vendor before renaming.
“What is eating my disk? List the twenty largest files.”
Walks your folders and explains what each large file actually is.
Writing and documents
Drafts, summaries and rewrites from material you already own.
“Draft a two-page proposal from the notes in Documents.”
Reads your notes, structures the argument, and types it into your editor.
“Summarize this PDF and pull out every date and amount.”
Opens the file locally and returns a table you can copy.
“Rewrite this letter in a firmer, more formal tone.”
Keeps your facts, changes only the register.
“Turn these meeting notes into an action list with owners.”
Names the owner and the deadline wherever the notes imply one.
Data and spreadsheets
The boring cleanup, done in one sentence.
“Merge these twelve CSV files and remove the duplicate rows.”
Matches on the columns that identify a row, not the whole line.
“Extract every email address in this folder into one clean column.”
Deduplicates, lowercases, and writes a new file.
“Convert this JSON export into a spreadsheet, one row per order.”
Flattens nested fields into readable columns.
“Check this ledger for rows where the totals do not add up.”
Recomputes each line and flags the mismatches.
Media and batch work
Hundreds of files in a single request.
“Resize every image in this folder to 1600px wide.”
Uses the tools already on your machine and keeps the names.
“Convert these thirty videos to audio only.”
Queues the batch, keeps the original files untouched.
“Rename my holiday photos by the day they were taken.”
Reads the camera date inside each file.
“Burn the subtitles from this file into the video.”
Runs the encode locally, then checks the result.
Code and the terminal
A capable pair of hands for scripts and small tools.
“Write a script that backs up Documents nightly and keeps a week.”
Writes it, runs it once, and shows you the output.
“Explain what this script does and where it could fail.”
Reads the file in place and walks the logic line by line.
“Add error handling to this module and run the tests.”
Edits the file and re-runs your test command.
“Watch this folder and rename new screenshots sensibly.”
Builds the watcher and starts it for you.
Thinking and planning
Reasoning over your own material, offline.
“Build a twelve-week study plan from the syllabus in Documents.”
Paces the readings across the weeks and marks the assessments.
“Compare these three quotes side by side and show what differs.”
Reads each PDF and returns a difference table.
“Turn this recipe list into a shopping list grouped by aisle.”
Merges repeated ingredients first.
“Draft a day-of timeline from the emails in this folder.”
Pulls the times, vendors and addresses it finds.
There is no command list.
The examples above are not the limits of the product. Halo reads your folders, opens your applications and writes to the screen, so anything you can describe — in your own words, about your own files — is in scope. Current local models are capable enough that the answer to “can it do this locally?” is usually yes. If it cannot, Halo tells you that, rather than quietly spending money to find out.
03 · When the cloud earns its keep
The cloud is the exception, and you approve it.
Four situations are worth a paid model. Before anything leaves your machine, Halo names the model, shows the rough cost and lists exactly what will be sent.
Live prices
Comparing stock and prices across shops needs the live web, so Halo asks before it connects.
Fresh research
News, recent releases and anything outside your files — with sources, not memory.
Very long documents
A book-length manuscript can exceed what a small local model holds in context.
The hardest reasoning
Novel strategy or subtle argument where a larger model genuinely earns its cost.
04 · Why Halo
Built to be trusted with your machine.
Stays above your work
A floating panel or a compact orb that sits over every window, including full-screen apps. Ctrl or Cmd + Shift + Space summons it.
Local by default
Runs on Ollama with Qwen, Llama or DeepSeek, and keeps those models current on their own.
Cloud on request
Escalates to Claude, GPT-5, DeepSeek or any of 100+ models through OpenRouter, using your own key.
Access you define
Minimal mode sees your applications and main folders. Full mode sees the whole disk. You choose at first launch.
Approval before change
Moves, writes and keystrokes can each require your confirmation, individually.
Every platform
macOS on Apple Silicon and Intel, Windows on x64 and ARM, Linux as AppImage, deb and rpm.
Try it on your own machine.
Free and open source. Install Halo, pull a local model, and give it the chore you have been putting off.