Build with agents.Show them what you mean.Stay in charge.
Mark up a document, a web page, or a photo, and your agents see exactly what you mean. Every model in one window, whether or not you read code.
Go ahead and click around. Answer the question, switch chats, type a message. Hover the numbers for a tour.
Everything in one place.
Other tools do one of these. Exarch does all three, with the same chats and documents.
An agentic development tool
An assistant
An agent dashboard
Exarch does all of it.
Mark it up. They get it.
Mark up a document, a web page, or an image, then say what you want. The agent sees your marks and makes the change. Try all three.
Every AI you use. One list.
Give the same request to three models at once and compare the answers. Split off a chat to try something, then bring the good parts back. Write your next message while it works, and it goes when the agent is ready.
Have one model check another.
Helpers are jobs your agent hands to a different AI. A review, a second opinion, a test run, a hunt for bugs. Claude can check GPT's work and the other way round, so nobody grades their own homework.
- Off until you turn it on, and you choose which models helpers may use.
- Ask for a model by name, or save a helper for jobs you repeat.
- Each helper starts on the account with the most room left.
- Different choices per project.
Helpers
Ask any available model to help, or save choices for recurring work.
Saved helpers
Your agent sees each helper's name and when to use it.
GPT-6.1-Sol → Claude Opus 5.5
Uses your general model order
GPT-6-Astra
Claude Fable 5.1
General model order
Used when a request has no exact model or saved helper.
The pickup form is built. Before we ship it, I asked for a review from a different AI.
The review found two things worth fixing, both in the pickup form. I've fixed them.
Helpers are off. Turn on Allow helpers in Settings, Helpers.
Two things worth fixing. The 4 PM slot can be booked twice when two people check out at once, and the confirmation email doesn't say where to collect the order. Everything else checks out.
Talk it through with Terra.
Say what you need out loud. Terra hands the real work to the agent in the chat, reads what it did, and tells you. Every word and every run is written into the chat, so nothing lives only in the air.
Voice runs on your ChatGPT sign-in. It isn’t private, so turning Private on stops it.
Terra remembers you.
Her name, her voice, and her memory are plain files in a folder you can open and edit. When you settle a chat, she writes up what she learned. What matters gets filed into her memory later, in the background, where you can read it.
- Rename her in IDENTITY.md. Change her voice in SOUL.md.
- Edit what she knows about you by hand.
- Nothing is remembered that you can't read.
Open the chat's ⋯ menu and choose Settle.
# User
How the owner prefers to work.
- Prefers short answers, decision first.
- Doesn't read code. Wants systems explained in plain words.
Go private with one click.
After a one-time setup, click the eye and your conversations go through a service that keeps nothing. Everything works the same, and that work stays apart from everything else.
Private is off. Your messages go to each AI the usual way, and everything here shows up in search. Click the eye in the sidebar.
The AI you already use
Claude, GPT, and others, sent through a service that promises in its contract to keep nothing. On the default Vercel AI Gateway, Exarch checks that promise once per session. On another host, it rests on your agreement with that service.
Locked-down hardware
Open models on machines that can prove nobody, not even the operator, can see your conversation. For work that can't leave a trace.
What stays out
Private chats, files, and memory never show up in your everyday chats or search. Private chats still get the built-in browser, and their screenshots stay in the private store. If the private connection fails, Exarch stops instead of quietly sending it the normal way.
Keep going from your phone.
Exarch Mobile picks up the same chats wherever you are. It only buzzes when an agent needs you or something went wrong. Finished work waits quietly.
- Send a message while the agent works, now or when it's done.
- Answer a question with a tap and get back to your day.
- Comment on a passage or mark up a photo.
Accounts, Library, schedules.
Keeps going when an account runs out
Sign in to more than one account. Exarch sticks with one while it has room. When it runs out mid-task, the work moves to another account on the same model and carries on by itself. Park one to keep it out.
One Library for skills and plugins
Skills teach your agents how you like things done. Turn each one on or off. Sharing your setup with your other computers is opt-in, one approved computer at a time.
Writes a plan you can read and decide on.
ClaudeCodexLogos, colors, and mockups in your brand.
ClaudeCodexGrokWhat surprised us, and what to do next time.
Schedules that land in a chat
Ask for something on a timer. Each run lands in a chat in the project you pick, either one chat that starts clean each run or a fresh chat every time. A desktop alert only if you want one.
Morning brief
Terra · Fable 5.1 · One chat, start clean each run
Check the stall's orders
Sunday Market · GPT-6.1-Sol · Fresh chat per run
Weekly newsletter draft
Sunday Market · Fable 5.1 · One chat, start clean each run
Also in the box: drag tabs to arrange your screen, pop chats and documents into their own windows, open your other computers in windows of their own with their chats, files and browser tabs, snooze or archive chats, search every chat at once, and pick a theme.
Nothing is held back.
You never have to touch these. If you write code, they're there from the first launch.
Worktrees
Each chat can work in its own copy of the project on its own branch, so agents don't trip over each other.
Pull requests
Link a PR to a chat. Checks and reviews sit beside the conversation, with a Fix button for anything that failed.
Terminal
A drawer below the chat, already in the right folder. Closing it keeps your session.
Permission modes
Supervised, Auto-accept edits, Auto, and Full access. Plan or Build, per chat.
Browser and devices
Agents drive a built-in browser and iOS and Android simulators, and show you the evidence.
Same tools for every agent
Every agent can edit documents, open tabs, and show its progress, whichever AI it runs on.
History import
Bring your existing Claude Code and Codex chats into a project with a short Import existing work wizard.
Every agent CLI
Claude Code, Codex, Grok, Cursor, and OpenCode run as they are, each with its own sign-in.
Open it up.
Linux, macOS, and Windows. And a phone app for the road.
Download and install
One package with everything included. The installer adds Exarch to your apps menu and sets up the exarch command.
Download for LinuxDownload the zip
The app isn't signed yet, so the first time you open it, right-click and choose Open.
Download for MacInstaller
A standard installer for 64-bit Windows. It isn't signed yet, so if Windows warns you, choose More info, then Run anyway. The first start takes about 20 seconds while it sets up.
Download for WindowsExarch Mobile
On the desktop, sign in to your Exarch account under Settings, Phone, and allow phone access. Then sign in on the phone and pick your computer.
Get the phone appThe questions everyone asks first.
Do I need to know how to code?
No. You describe what you want, agents build it, and everything that needs you comes back as a document, a question, or an approval. If you do write code, the developer tools are there from the first launch.
Which models does it work with?
Claude, GPT, Grok, Cursor, and OpenCode, which opens the door to dozens more. Sign in to as many accounts as you have. Exarch sticks with one account while it has room, then moves to the one with the most left. Or you choose.
Does anything leave my computer?
Your messages go to the AI you chose, the way they would from that company's own app. Documents, chats, memory, and settings stay on your machine. There's no Exarch cloud storage and no telemetry. Once a day Exarch asks, anonymously, whether a newer release is out. If you turn on phone access or share your setup with another computer, that traffic passes through an optional Exarch account relay, end-to-end encrypted and only in transit.
What's the phone account for?
The desktop app needs no account. Phone access signs in to an Exarch account on both the computer and the phone, so Exarch can prove it's really your phone and connect it safely to your computers. Private opens on the phone only after you approve it on the computer.
What does private mode actually promise?
That your private conversations go only through a service that keeps nothing, and stay out of everything else. The Private section covers the two ways to run it.
What does it cost?
Exarch is free, with every feature included. You pay for your AI subscriptions as you already do. Optional paid services, like setup and hosted private servers, are coming later.
How do updates work?
Exarch checks for a new release when it starts and about once a day. When one is out, a single line tells you the next step: one command on Linux, a download on Mac and Windows. Nothing downloads or installs on its own, and you can turn the check off under Settings, Advanced.
Who makes it?
Exarch is built by someone who works with agents every day without writing code, for everyone else who does the same.
Your agents are waiting.
Download Exarch, add a folder, and start a chat. Everything that needs you comes back to you, ready to mark up.
Download