What is Halo? Computer Use Without an LLM

Halo is a novel neural, deterministic runtime for computer use. You ask for a task in plain English, like "find the 10 largest PDFs in Downloads" or "what's using port 8080", and Halo turns it into exact, predictable steps that run on your own machine. It does not use a large language model to do it, which makes it fast, inexpensive to run and safe by design.

What Halo is

Halo understands what you want and turns it into exact, supported computer actions. A request can be a single action or several steps that build on each other, all within a defined set of computer tasks: files, search, data, archives, media, processes, networking, Git, packages, scheduling and more. It shows you how it understood you, and it asks before it changes anything.

Three things set it apart:

• Deterministic. For a given Halo model, the same request in the same relevant context and machine state gets the same plan. A Halo model ID always behaves the same way, and Halo never silently switches it.
• Inexpensive to run. There is no large language model behind each request, so there is no per-token bill to pass on to you.
• Bounded. Halo can only carry out actions it was built to perform, and it enforces the permissions required to execute them. It cannot invent unsupported actions or silently step outside its capabilities.

Halo is in Research, runs on macOS and Linux, and is free during Research.

What Halo is not

• It is not an LLM or a chatbot. Halo does not generate text or code to answer you. It understands your request and turns it into exact, supported actions.
• It is not an autonomous agent. Halo can carry out bounded multi-step jobs, but it does not invent goals or roam through your system. It does what you asked, after you see it.
• It is not a coding agent. Halo is built for computer tasks, not for writing code. For writing code, Memcode has a separate coding agent.
• It does not write arbitrary scripts. When something is outside what Halo does, it says so instead of guessing.
• It is not a general assistant. Ask it to book a flight and it will tell you that is not something Halo does.
• It is not a bulk automation backend. Usage limits keep the service fast for people using it interactively.
• It is not on Windows yet. Halo runs on macOS and Linux during Research.

You can always run a command yourself. Start a line with $ and Halo runs it directly as a shell command.

The problem Halo solves

Most of us know what we want the computer to do. The hard part is remembering how to say it in shell.

Which flag sorts by size? What is the syntax to find files modified yesterday? How do you see which process is holding a port, then stop it? You search, copy a command from a forum, tweak it, and hope it does what you think.

The usual fix today is to ask an AI chatbot or agent. That works, but you pay for every request, you wait for the answer, and you get a command that might be wrong, so you still have to read it carefully before you run it.

Halo removes both steps. You say what you want in your own words, and Halo runs the right steps, quickly, at a fraction of the cost.

◯ find the 10 largest PDFs in Downloads
◯ what's using port 8080
◯ find the PDFs from yesterday and zip them

There is nothing to memorize, and you can phrase requests your own way. If Halo is unsure what you meant, it asks.

How Halo is different from using an LLM

Large language models are remarkable at open-ended work like writing and coding. For everyday computer tasks, that open-endedness is mostly cost and risk. Halo is built the other way around.

Cost. This is the biggest difference. An LLM is expensive to run, and that cost lands on you as a subscription, a token bill or usage caps. Halo does not run an LLM for your requests, so the work costs a small fraction as much. That is why Halo can be free during Research and why it can stay affordable for the hundreds of small tasks in a normal week.

Safety. An LLM agent can write any command, including a wrong or destructive one, and it can chain steps you did not expect. Halo can only do what it was built to do, enforces the permissions each action needs, shows you exactly what it will touch, and waits for your yes before changing anything.

Predictability. Ask the same thing twice in the same relevant state and context, and Halo makes the same plan twice. When the state is different, for example when other build files exist, the plan reflects what is actually there. That makes it something you can rely on, and something your own programs can build on.

What Halo can do

Halo handles a defined set of everyday computer tasks, and that set grows throughout Research. Here is what it covers today:

Inside Halo, type /capabilities to see the same list with examples.

Everyday file chores

Find, sort, count, move, copy, rename and delete files. Ask for "PDFs in Downloads from last week", "files bigger than 100 MB, not in node_modules", or "move the screenshots from the desktop into a folder called shots". You can also open a file, such as "open the latest PDF in Downloads".

Data, documents and media

Query CSV and JSON files and convert between them: "the 5 biggest orders in orders.csv" or "convert every CSV in exports to JSON". Read the text of a PDF, merge invoices into one PDF, resize photos, or extract the audio from a video. Media conversion uses tools already on your machine, such as ffmpeg, and Halo tells you when one is missing.

Developer and system work

Halo is also built for the work around code: Git, Docker and Compose, processes and ports, logs, Python and Node projects, builds and tests, packages and runtimes, and databases like Postgres and Redis. A few examples from the Halo page:

• "why won't postgres start?" checks the service, its port and its logs, then gives you the finding.
• "run the tests", then "rerun the failed ones", reruns the same failures.
• "show errors from the last hour in worker.log" pulls the matching entries.
• "what node version does this project expect?" compares the project's setting with what you have installed.

Follow-ups know what came before, so you can keep a conversation going instead of starting over each time.

Multi-step jobs

Some requests have several parts. "Find the CSV files from this week and convert them to JSON" is two steps. Halo shows the whole job before anything runs, and one approval covers all of it. Later steps use exactly what the earlier step found, not a fresh search. If a step fails, the rest does not run, and Halo shows you where it stopped.

Scheduling

Ask "every weekday at 9 run sync.sh" and Halo shows the schedule it understood before creating anything. You can list your scheduled jobs and remove them the same way.

What are the risks?

Halo is new, experimental technology. It is designed to be bounded and predictable: it only works through supported capabilities, shows changes before they happen, and asks for your approval before modifying your machine. Those safeguards reduce risk. They do not remove it.

Halo can misunderstand a request. Software can contain bugs. And an action that is correct on its own can still have unexpected results because of the state of your machine or the tools installed on it.

During Research, we recommend against using Halo on:

• production systems;
• machines where downtime or lost data would be costly;
• systems that hold sensitive, confidential or regulated data;
• critical infrastructure or business-critical environments;
• anywhere an incorrect action could cause serious loss or disruption.

Halo Research is a good fit for hobbyists, developers, researchers, technical enthusiasts and anyone interested in exploring new approaches to AI and computer use, who understands they are trying experimental software and is comfortable reviewing what it proposes.

If that is you, a few habits go a long way. Read what Halo proposes before you approve a change. Keep backups of anything important. Use it on machines and folders where a mistake is recoverable.

You stay in control

Halo's rule is simple: it won't change your machine without showing you what it plans to do and getting your approval.

Read-only actions run straight away. Listing files, searching, showing processes, Git history, system information and calculations change nothing, so there is nothing to approve.

Changes wait for you. Before anything is changed, you see what it affects:

A few more details that matter in practice:

• Deletes go to the Trash unless you ask for a permanent delete.
• Exactly what you approved. Halo acts on exactly the files it showed you. Files that start matching afterwards are not added. If something changes after the preview, Halo stops.
• Your working folder is the default. Halo works in the folder you start it in. Reaching beyond that folder, such as other folders, processes, network ports or system information, needs your permission.
• Some permissions are never remembered. Administrator access and stopping processes always ask, every time. /permissions shows what you have allowed.
• Even $ commands get checks. Risky commands ask first, and some commands are always blocked.

Why it saves you money and time

It costs far less to run. Halo does not send your requests to a large language model, so there is no per-token cost behind each task. That is the main reason Halo exists: everyday computer use should not need an expensive model. Halo is free during Research, and pricing after Research will be announced before it changes.

It saves your time. Tasks you would normally look up, test and fix run in the time it takes to type a sentence, without waiting on a chat to finish writing.

It is one simple tool. Halo is one program for macOS and Linux that runs in the terminal you already have. There is no heavyweight app to set up for chores like cleaning a folder or finding a stuck process.

Halo Research has these usage limits per account:

Limits roll, so capacity frees up gradually rather than at a fixed reset time. Normal use should almost never reach them, and if you do, Halo tells you how long to wait. Commands you run directly with $ never count against the limits.

Use it from your own programs

Halo also has an Agent SDK, starting with Go, with more languages to follow. Your programs get the same checks and approvals as the interactive shell, with results they can act on. See the Agent SDK docs.

Frequently asked questions

Does Halo use an LLM?

No. Halo is a neural, deterministic runtime, not a large language model. It understands your request and turns it into exact, supported actions, which is why it is fast, predictable and inexpensive to run.

Why is Halo cheaper than an AI agent?

AI agents send every request to a large language model, and you pay for that in tokens, subscriptions or usage caps. Halo does not run an LLM for your requests, so each task costs a small fraction as much. It is free during Research.

Is Halo safe to use on my machine?

Halo can only perform actions it was built to perform, so it cannot invent unsupported ones, and it enforces the permissions each action needs. Anything that changes your machine shows what it affects and waits for your yes, and stopping processes or using administrator access asks every time.

Should I use Halo on a production machine?

No, not during Research. Halo is experimental technology, and we recommend against using it on production, business-critical, sensitive-data or otherwise high-consequence systems. See What are the risks? above.

Is Halo free?

Yes, during Research. Reasonable usage limits apply, and pricing after Research will be announced before it changes.

How is Halo different from an AI coding agent?

Halo is built for computer tasks such as files, processes, Git, archives and media, not for writing code. For writing code, use Memcode Code.

Which operating systems does Halo support?

macOS and Linux, on x86_64 and arm64. Windows is not supported yet.

Do I need an account?

Yes, Halo uses your Memcode account. If you already use the Memcode CLI, you are already signed in, because the two share one sign-in.

How to get started

Install Halo with one line. It does not need administrator rights:

curl -fsSL https://memcode.ai/halo/install.sh | sh

You can read the install script first if you prefer. Then start Halo and sign in:

halo
halo login

halo login opens your browser to sign in with your Memcode account. After that, just ask:

◯ find the 10 largest files here

For a single request without opening the shell, run halo "what's using port 3000". Keep Halo current with halo upgrade.

Where to go next:

• Halo product page
• Quickstart: your first request and your first approval in five minutes
• Installation
• Supported tasks
• Safety
• Usage limits
• Privacy