Chat The harness How it works Stack Can and can't Pricing About Request trial
Private trials Public beta Q4 2027

An agent harness that runs on the machine you already use.

computer.vodka is an ai agent harness. It is the control loop around the model you already run. Each turn is the same: plan, pick one tool, read the result, then stop or ask before a step that is hard to undo. Tools, memory, and approvals stay in one folder on your disk. The install is free. Hosted extras are optional and metered.

Public beta opens Q4 2027, after private-trial bugs are actually fixed. Thanks for waiting.

computer.vodka · session probe

You type the task. The harness runs the loop.

A harness is the runtime around a model: planner, tool runner, workspace root, session log, and a permission gate. You talk here. It proposes a plan, calls one tool, reads the output, and waits if the next step would delete, overwrite, install, or send. Quality follows the weights you point it at. This page is a preview of that chat shell. The real loop starts after you install.

computer.vodka
Online

These are the connectors the harness can switch on inside the chat. Hover or tap the row to pause it. Live OAuth ships with the trial. Hosted extras keep them alive when the laptop is closed.

Not a solo agent. The loop around any model you run.

Most products hide the model and ship your files to a GPU farm. A harness does the opposite. It is a small process on your machine: a planner, a strict tool schema, markdown memory, and a gate in front of writes. You attach Llama, Qwen, Mistral, Gemma, or an OpenAI-compatible server on localhost or your LAN. Swap the weights and the product does not change. Free local quality is exactly as good as those weights. The subscription only covers work a closed laptop cannot do alone, like always-on text and multi-device sync.

A model-agnostic agent loop

Plan, pick one tool, observe, repeat. Talks to Ollama, llama.cpp, LM Studio, vLLM, or any OpenAI-compatible /v1/chat/completions endpoint on localhost or your LAN. Swap the weights without rewriting the product.

Your files stay put

Every path is resolved against one workspace folder. Memory is markdown you can open. The installed harness does not phone home with prompts or file contents.

Trimmed for a laptop

Rolling summaries, a small tool schema, and interruptible streams so a long session still fits on Apple Silicon or a strong CPU.

Approval before damage

Deletes, overwrites, installs, and outbound sends wait for a yes. You see the plan and the diff before anything hits disk.

Tools, not a plugin zoo

read, write, search, shell, fetch, diff, plus connectors you toggle in the chat. Enough to finish work. Not a kitchen sink.

Apps as switches

Mail, calendar, docs, chat, and the rest sit behind the Apps toggle. Off means the tool is never offered to the model.

Text the harness

Give it the number you want watched. SMS and WhatsApp then drop a task into the same loop. The first message in a thread is password-checked so a random sender cannot act for you.

Free local, hosted extras on a sub

The loop on your disk is free and follows your model. Cloud fallback, multi-device sync, always-on text, and scheduled jobs are usage-billed.

Three steps, then the harness lives on your desk.

No account to babysit for the local path. You go from a trial request to a working loop in one folder. After that, every turn is plan, one tool call, observe, then continue or stop. Quality is the model you attach. Hosted extras stay optional.

1

Ask the chat for a trial

Open the interface below and tell the agent you want in, or just type your email. It collects what it needs and sends the request for you.

2

Install the harness

macOS, Linux, or Windows. Point it at Ollama, llama.cpp, LM Studio, or any OpenAI-compatible endpoint. After setup the local loop does not need an account.

3

Type the task

Point it at a folder, say what you want in plain language, and let the loop plan. Approve writes, deletes, and installs. Read the transcript after.

A local runtime, split into parts you can inspect.

The site is a preview. The installed product is one process: planner, tool runner, memory on disk, policy gate. That is the whole idea of a harness. It does not pretend to be the model. It wraps whatever you attach, keeps paths inside one workspace root, and writes a plain transcript of every tool call. Free local quality tracks those weights. Subscription extras are only the hosted pieces a single disk cannot cover.

The loop

Every turn is the same four steps: plan, pick one tool, read the result, then decide whether to continue or stop. Context is trimmed each turn so a long session still fits on a laptop GPU or a strong CPU.

  • 01Talks to Ollama, llama.cpp, LM Studio, or vLLM over any OpenAI-style /v1/chat/completions endpoint on localhost or your LAN.
  • 02Model-agnostic: Llama, Qwen, Mistral, or Gemma, and whatever you already run. Quality follows the weights, not a hidden API.
  • 03Tool calls use strict JSON schemas. A malformed call is caught and retried instead of crashing the turn.
  • 04Context window is managed with a rolling summary, so multi-step tasks stay in one transcript without overflowing the model.
  • 05Streaming tokens render live, and any turn can be interrupted mid-run without corrupting the workspace.
  • 05bOptional hosted router on a subscription: if the local model stalls on a hard turn, that one step can bounce to a cloud model. Files stay in the workspace unless you attach them.
  • 05cSystem prompt and tool list are generated from the capability switches, so a disabled app is invisible to the model, not just ignored.

Tools

Enough to finish a job in a folder, not a kitchen-sink plugin list. Risky calls wait for a yes.

readwritesearchshellfetchdiffconnectmailcalendarsms
  • 06Workspace is a single folder. Every path is resolved and checked against that root, so it cannot walk your home directory.
  • 07Writes go through a diff preview. Deletes, overwrites, and install commands pause until you confirm the plan.
  • 08Shell runs with a timeout and captured stdout/stderr, so a hung command can't lock the session.
  • 08bFetch is allow-listed to URLs you ask for. It is a page reader, not a crawler. It does not wander the open web on its own.
  • 08cConnectors speak scoped OAuth. Tokens live in a local key store. Revoke per app without touching the rest of the harness.

Memory

No hidden vector store you cannot open. Notes live as plain markdown in the same workspace.

  • 09Memory is a set of .md files you can open, edit, or delete like any other file.
  • 10Optional local embeddings index the workspace for search. It stays on disk and never leaves the machine.
  • 11Every tool call is appended to a plain-text session log, so you can audit exactly what ran, in order.
  • 11bNotes can be pinned into the next turn as working memory. Nothing is buried in a vendor vector store you cannot open.

What stays local

Prompts, file contents, shell output, and embeddings stay on the box you run. This page only emails a trial request. The installed agent does not phone home.

  • 12A single binary plus your model. macOS, Linux, and Windows, first-class on Apple Silicon, which is what we develop on.
  • 13No telemetry and no account after setup. You can run it fully air-gapped once the model is pulled.
  • 14Config is one readable file. No cloud project, no API key required for the local path.
  • 14bFirst-class on Apple Silicon, which is what we develop on. CUDA and CPU backends through the model server you already chose.

Connectors and text-in

Apps in the chat are switches, not a second control panel. Live connections ship with the trial. You can also text the agent once a conversation is unlocked.

  • 15Mail, calendar, docs, and chat apps sit behind the Apps toggle. Off means the tool is never even offered to the model.
  • 16Connectors use scoped OAuth. Tokens are stored locally and you revoke access per app whenever you want.
  • 17The first message in any text thread is password-checked, so a random number cannot act on your behalf.

What the harness can do, and what it cannot yet.

Type a task and the loop tries it: plan, tool, observe, stop. How good the answer is depends on the model you attached, because the harness is not a secret smarter model. Free local follows those weights. The subscription adds hosted reasoning, sync, and always-on text. Video generation and native video edit are the clear not-yet. Almost everything else is already in the runtime.

It can

  • Read and edit a workspace folder

    List, search, open, rewrite, and grep across one directory you choose. Paths cannot walk outside that root.

  • Write drafts, data, and scripts

    Markdown, JSON, CSV, HTML, Python, R, shell, SQL. Diff preview before anything stays on disk.

  • Run local commands you approve

    Shell, tests, formatters, builds, short programs. Installs and deletes wait for a yes.

  • Wrap any model you already run

    Ollama, llama.cpp, LM Studio, vLLM, or any OpenAI-compatible server. The harness is the same. The brain is yours.

  • Keep a readable session log

    Every tool call lands in a local transcript. Open it like any other file. Audit what ran, in order.

  • Keep working offline after setup

    Once the binary and model are on the machine, prompts and files do not go to our servers.

  • Talk to connected apps

    Mail, calendar, docs, chat, drive, and the rest of the ticker below the chat. Off means the tool is never offered.

  • Pick up a task from a text

    SMS or WhatsApp, after a password check on the first message. Hosted gateway is a subscription extra so it works when the lid is closed.

  • Research, summarise, and extract

    Open a page you name, pull quotes, turn them into notes. With local files on: grind a folder of PDFs, CSVs, or code inside the workspace root.

  • Code, debug, and test in the folder

    Read a repo, patch a file, run the test command you approve, paste the failure back into the next turn.

  • Draft mail, briefs, and reports

    Write the thing, show it, wait for send. Connected inboxes are a connector, not a default outbound pipe.

  • Images the model already supports

    Describe, caption, and file-handle images in the workspace. Generate stills only if the attached model can. That is the model, not a hidden API.

  • Tables, spreadsheets, and light analysis

    Read CSV or Excel-like files, compute, chart to a file, write a summary. Heavy stats still want a model that can count.

  • Remember across turns on disk

    Plain markdown notes plus an optional local embeddings index. You can delete a memory by deleting the file.

It can't, at least not yet

  • Generate or edit video

    No native video model in the harness yet. It can organise clips on disk and write ffmpeg command plans you approve. The pixels themselves are a later problem.

  • Match frontier cloud models on the free path

    A 7B or 14B on a laptop is not GPT-class reasoning. Subscription hosted extras exist for the hard turns. Files stay local unless you attach them.

  • Run well on a weak machine

    Useful local models need RAM and, ideally, a GPU. Older phones and low-RAM laptops will stall.

  • Browse the open web by default

    Network fetch is off until you turn it on, and even then it is for pages you ask it to open, not a general crawler.

  • Always read the task correctly

    It can mis-scope a folder, skip a file, or write the wrong thing. Review the diff before you approve.

  • Replace git or backups

    There is no built-in version control. Keep your own copies of anything you cannot afford to lose.

  • Act from this website

    The chat on this page is a preview. It cannot see your disk. The loop starts after you install the trial.

  • Be a single branded brain

    computer.vodka is the harness. It is not "the model." If the weights are thin, the answers are thin. That is the deal on the free path.

The harness stays free. Hosted extras are usage-only.

Install it, point it at your model, work in a folder. No card. The free path is the full local loop: plan, tool, observe, approve. How well that works is how well your local or LAN model works. From the Q4 2027 public beta, hosted extras are optional and metered. Cloud reasoning for hard turns, multi-device sync, always-on text, live connectors, jobs that outlive sleep. No fixed seat fee. Quiet months stay cheap. Anything that can run on your machine stays in the local product.

Always free

Stays local

No subscription · ships with the trial · quality = your model

  • The full harness loop

    Plan, tool, observe, stop. Read, write, search, shell, and local fetch inside the folder you choose.

  • Your own model

    Ollama, llama.cpp, LM Studio, or any OpenAI-compatible endpoint on the machine or LAN.

  • Memory and logs on disk

    Markdown notes and a plain transcript. You can open or delete them without an account.

  • Approval before damage

    Deletes, overwrites, and installs wait for a yes. That gate does not move to the cloud.

  • Works offline after setup

    No seat, no monthly fee, no phone-home just to keep using the agent on one computer.

Q4 2027 · pay as you go

Hosted extras

Usage billed each month · no fixed fee, no send limit, only what you use

  • Cloud model when local is not enough

    Metered hosted reasoning for the hard turns. Files stay in your workspace unless you attach them.

  • Same session on more than one device

    Sync chat and memory notes across a laptop and a phone. A single disk cannot do that by itself.

  • Always-on text and WhatsApp

    A hosted gateway so the agent can pick up SMS or chat when the laptop is closed. Bill by messages handled, not a plan cap.

  • Live app connectors

    Gmail, Calendar, Slack, Drive, and similar need an OAuth broker we host. You pay for the calls you make.

  • Jobs that outlive sleep

    A hosted runner keeps a long task going after the lid closes. Local processes die with the machine.

  • Team approvals and shared logs

    A second person can approve a risky step and read the same transcript. That is a server product.

  • Usage dashboard and receipts

    See tokens, connector calls, and messages for the month. Export a receipt. Stop anytime and local keeps working.

  • Scheduled jobs and webhooks

    Wake the agent on a timer or an inbound event. Hosted only, because a closed laptop cannot keep a clock.

Four people who wanted the harness on their own laptop.

We design, build, and run the harness on our own machines in Aotearoa. Raman is in Wellington. Jayson, Ashly, and Emily are in Auckland. Each person owns a part of the runtime (loop, evals, interface, safety) and still reviews the rest.

R

Raman

Runtime and inferenceWellington

Owns the harness loop: plan, tool call, observe, stop. Makes local models through Ollama, llama.cpp, and LM Studio feel usable on a laptop, especially Apple Silicon. Also keeps the tool runner and context trimming from grinding a long session to a halt.

J

Jayson

Product and evaluationAuckland

Turns raw agent behavior into a task you can hand someone who does not live in a terminal. Writes evals against folders, decides what ships in a trial, and cuts anything that only looks good in a demo. Watches private-trial bugs so the Q4 2027 public beta is actually usable.

A

Ashly

Design and interfaceAuckland

Designs the chat, the capability switches, and the approval states. Builds the desktop shell so computer.vodka feels like a modern tool instead of a raw CLI. Also owns the small details: the mark, the trial flow in the preview, and how the product reads on a phone.

E

Emily

Memory and tool safetyAuckland

Works on local notes, session logs, and the permission rules around write, delete, and install. Makes sure a bad plan cannot walk outside the folder you pointed at. If a step is hard to undo, it waits for a yes before it runs.