What changed, and when

Computer use. Atlarix can now see your screen and use the mouse and keyboard, on macOS, Windows and

Linux. It's off until you switch it on for a model, in Settings, the chat's model menu, the terminal or your

editor, and only models that can see images can use it. While it runs the app shows "Controlling your

screen" with Stop; moving the mouse means its next click or keystroke is refused, and Esc stops it at once.

It works on whatever is open, so close what you don't want it to touch.

Sessions read each other's results. One session can see what another just finished: how it ended, its

reply and the files it changed.

Clickable links. Links are clickable in the terminal, in your editor and in the VS Code panel, including

the addresses your dev server prints.

Lighter. The phone tools, the first-run tour and the rotating tips are gone; the agent drives a phone

emulator with ordinary commands and screenshots, and models that can't see images aren't offered image tools.

Chrome extension. "Allow on every site" lets Atlarix work on any site without asking site by site;

password fills and other sensitive actions still ask.

The cloud agent is a paid feature. Cloud runs from Slack or atlarix.dev use Atlarix's models and your

Atlarix balance; use your own key in the app, the terminal or your editor.

Sessions steering sessions. One session can set another session's mode and model, and review its plan:

approve it to start building, send it back with a reason, or replace it. The other session always shows who

changed what.

How is Atlarix doing? After a few turns Atlarix asks once to rate the session 1–3, and /feedback sends

a note any time, in the app, the terminal and your editor. Nothing is sent unless you answer. The "where did

you hear about us" question is gone.

Session names follow you. A chat's name shows on its VS Code tab and header, and a rename made anywhere

reaches every open window within a second.

Searches that find everything. File search and text search now reach folders outside the project, ~

paths, single files and dot-folders, and respect each folder's own ignore rules.

Atlarix in Slack. Reply in a task's thread to keep working on the same pull request, ask a question

without changing any code, or ask for a review.

Also fixed: VS Code saves a new chat only once you write in it and says why when it can't start; a

repository's settings can no longer choose which program the panel runs; atlarix run explains when a

custom endpoint needs a key; the app keeps working through a short outage of Atlarix Core.

Project memory. Atlarix keeps what it learns about a project as short notes. A new session starts with

the list of them, each message brings along the few notes that matter to it, and Atlarix notices when the

notes change elsewhere. When the list grows long, it asks to tidy it up.

Live sessions. A session open in two places stays in step. When the CLI, the app, VS Code or Zed is

running a turn, the others follow it as it happens, with Stop; what you type there goes to the running

turn. `/fork` copies a conversation into a new session in the CLI, the app, VS Code and editors that

support it.

Approvals. After you answer a card, one line says what you allowed. A rule you save ("allow edits in

src/") answers the cards it covers at once, and nothing runs twice. Risky git commands are caught however

they are written.

Compaction. /compact counts what your tools returned, summarises with the session's own model and

says when a summary fails. A request too long for the model is compacted and retried instead of failing.

MCP servers. /mcp in the terminal says whether each server connects, and why not when it doesn't. A

server's error output no longer prints over the terminal. In the app, the server list follows changes made

in the terminal, switches respond at once, and an automation's tool is picked from the server's own tools.

Also fixed: several Atlarix on one machine no longer mix up sessions, approvals or temp files; a turn

stopped from another surface ends as stopped in your editor; deep thinking asks for the model's highest

level; web sources show as links; VS Code's panel has copy buttons on code blocks.

What Atlarix counts. Analytics (only if you said yes) now say which kind of provider and model a turn ran

on, never a provider's name or address, and never your prompts, code or keys. A turn that fails on your own

key or local model sends a crash report with the kind of failure only. Signed-in accounts keep a daily count

of their turns by surface and model, as Atlarix already did for Core and Auto; the consent question and the

privacy page say so.

Every request is lighter. Today's models need to know what is true and what is allowed,

not how to work, so the fixed part Atlarix sends with each message is about a third smaller,

with every tool still there. Rules are said once, the coaching is gone, and the ones that

prevent real mistakes stay.

Browser: reading a page's console right after opening it now includes what the page

logged while it loaded. A tab that is already closed reports closed instead of refusing.

Also fixed: DeepSeek through Amazon Bedrock no longer shows a fragment of its tool-call

markup in replies.

Thinking works on the newest Claude models. Sonnet 5 and 5.5, Opus 4.7 onwards, Haiku 5.5 and

Fable refused requests with thinking on, whether through your own Anthropic key, Amazon Bedrock or a

gateway. They now think adaptively, and /effort sets how hard.

A fifth Core model. Core can now offer GPT-6 Luna beside GLM-5.3-Flash and the rest of the lineup.

Web search in the terminal and your editor. On a paid Core model or paid Atlarix Auto, search

now runs through Atlarix and is billed per search to your balance. The desktop app keeps searching

with its own built-in browser for everyone, at no cost. If your balance can't cover a search, you're

told so instead of getting an empty result.

Also fixed: a Core model picked in a newer Atlarix is no longer answered by a different model in

an older one; you're asked to update instead. The terminal's model list shows Auto as "Free" only on

a free account, shows each Core model's real context window, and names the model under each reply.

Microsoft Store users get new versions sooner.

Bring an endpoint that speaks Anthropic's API. When you add a provider of your own, choose

OpenAI-compatible or Anthropic, in the app, the CLI, your editor and VS Code. That covers Amazon

Bedrock's Anthropic endpoint, company gateways and proxies. Atlarix works out how the endpoint wants

your key and never saves one it refuses. Thinking, effort levels and prompt caching work as they do

on Claude directly.

Build mode reuses its prompt cache. Each message's own words were sent inside the part the

provider caches, so every message paid to cache the whole prompt again. From the second message on,

the fixed part is now read from the cache, at a small fraction of the cost.

`atlarix` opens at once. It used to show nothing for about two seconds while it looked up every

model from every provider just to name yours in the header. It now remembers the name and draws

straight away, then checks the list in the background.

Also fixed: effort "low" on Claude and other budget models now thinks less than the default

(it used to match it), and a key a provider refuses no longer replaces the one you had.

The app starts lighter. It checks your account once when it opens, instead of once for each

part of the window that shows it, and reads your plan in one request instead of several. Upgrades

still show up as soon as you switch back to the app.

The MCP marketplace in the CLI and VS Code now says how many servers there are (or how many

match your search) and lets you page through all of them. They used to stop after the first page.

Signed out, `atlarix` opens on a welcome screen with the two ways in: your browser (GitHub,

Google or email), or a code for SSH and machines with no browser. Nothing opens until you choose,

and the screen shows each step of the sign-in.

/effort replaces deep thinking with one control for how hard the model thinks: in the app's

chat menu, /effort in the CLI (ctrl+t for the top level), and in VS Code, Cursor, Zed and JetBrains.

On your own key or a Core model you get that model's own levels, exactly as it lists them, and the

one you pick is sent unchanged. On paid Atlarix Auto the levels choose the models: low uses the

small, fast ones, medium is the usual Auto, and max the strongest. Free Auto runs at one level.

Also fixed: a reply that asks you a question keeps what it wrote before the question once you

answer; in the CLI, background "finished" lines stay where they happened, long commands no longer

hide the Yes / No choices, and the tab says <!> needs you when Atlarix is waiting on you; choosing

"none" or "off" really turns thinking off, and Gemini gets the level you pick.

Your activity. See how you use Atlarix on this computer: lifetime tokens, your busiest day and

longest chat, streaks, a year-long activity calendar, your favourite models and modes, and the

skills and MCP servers you use most. On the Account page in the app, in VS Code's Account panel, and

as /stats in the CLI and any ACP editor. It's stored only on this computer.

More free Atlarix Auto. Free Auto can now answer from more free models. Because that shares your

prompts with more providers, Atlarix asks you once more before you use it: two short lines, and VS

Code, Zed and JetBrains can show it too.

Choose the MCP servers every workspace offers. Star a server in the marketplace (or ask the AI)

and it shows up, ready to connect, in the chat menu of every workspace, new ones included.

The Chrome extension asks before every call to a connected service, so a web page can never get

the agent to act on your accounts without your yes, and it only asks for a call that can run.

Also fixed: a stopped reply stays stopped (marked right away, kept after you quit, never re-run

when you send the next message); switching workspace from the ⌘K search opens that workspace's own

chat; edit cards show the real change; without Git installed the chat says Revert is off; search

results show a chat's words, not its markup; chats on free Auto get proper titles; removing a

server that's on all your devices asks first.

The marketplace is the official MCP Registry, with Sentry, Notion, Supabase, Canva and Linear

Verified on top and added in one click with a browser sign-in. Every card shows who published it,

whether it runs on the web or on this computer, and its logo. It works offline too.

Web servers follow you everywhere: one at a public https address is saved to your account, so

the app, the CLI, your editor and the Chrome extension all have it. Paste a URL, a command or a

mcpServers block from another tool and each server goes where it belongs.

Atlarix finds the server a task needs and offers to connect it, and your editor's own MCP

servers now work in Atlarix's chats there.

Also: Revert keeps the changes you made after a turn and names them; automations can watch your

account's MCP servers; messages from another Atlarix session read "from contributor".

A finished automation wakes the agent in its folder, in the app, the CLI, your editor or any ACP

editor, and the agent acts on the result. Automations have their own page in the editor, with Run

now, on/off and delete.

Every provider and its models, before you connect one. Models & keys lists every provider with

its models; connect any OpenAI-compatible address, and a local server such as Ollama lists its own

models. A provider that can't be reached says so, and says when it's retrying.

The CLI and the editor panel read like the app: new colours, thoughts, tables, diagrams and

code drawn properly, clearer plans with Build where the plan ends, and messages typed mid-turn that

steer or queue and are never lost.

Also: on Windows, PowerShell downloads and deletes always ask, like curl and rm; the editor's

sign-in and Account view are redesigned; atlarix --help lists the commands.

Atlarix in VS Code, Cursor, Kiro and Windsurf, rebuilt. A bar across the top for a new chat,

past chats, MCP servers, your account and settings. A task card shows the chat's tokens, its cost

on your own keys, how full the context is, and the plan ticking off as it is built. Edits and

commands arrive as cards; after a turn that changed files, Compare shows each file then against now

and Restore goes back. Approvals wait above the message box and say why a command is risky. Type @

for files, folders, Problems, the terminal, your git changes or a page, and / for commands and your

skills. Go back to any of your messages and send it again. Past chats has search, every folder,

stars, rename and delete.

Your own keys, managed properly. Every key is checked with its provider, before it's saved too;

a rejected key says so and offers Replace. Each provider has a page with its models' prices and what

they can do. Switch a provider off without losing its key, add any OpenAI-compatible endpoint, star

your favourite models, and optionally plan with one model and build with another.

Your MCP servers, found. Servers you set up in Claude Code, Cursor, VS Code, Kiro, Windsurf, Zed

or Claude Desktop appear in Atlarix with an Add button, in the app, the CLI and your editor, and

added ones are shared by all of them.

Zed and JetBrains. /keys shows your keys and whether each works, atlarix keys add adds one

in a terminal, /mcp add brings in a server you already have, and your skills are commands.

Also: "Edit files in this folder without asking" in Auto-approve, and saving a key no longer

loses other providers' settings.

After you've used the Atlarix extension for a few chats, it asks once whether you'd like to rate it,

and never asks again. Ratings are how other developers find it in the Marketplace and Open VSX.

The Atlarix extension's pages on the VS Code Marketplace and Open VSX now link to its repository

and show that it's free. Nothing changes in the app, the terminal or the extension itself.

Install Atlarix in VS Code, Cursor, Kiro or Windsurf. Search for "Atlarix" in the VS Code

Marketplace or Open VSX; the Atlarix CLI comes inside the extension, so there's nothing else to

set up. You get everything you use in the app: Build, Explore, Plan, Debug and Review modes; any

model, including your own keys; edits shown in the editor's own diff before anything is written;

plans you can read and build with one click; questions answered from a card; your Chrome; and

messages between your Atlarix sessions. Your account is there too: plan, credit and usage, Pro or

credit, managing or cancelling your subscription, signing out, your own API keys, and this

folder's MCP servers. Run out, and the message offers credit, Pro or your own key right there.

Take a chat anywhere. Type /editor in the terminal, or pick "Open in your editor" on a

session in the app, and the chat opens in VS Code, Cursor or Kiro. From the editor you can carry

it on in the app or the terminal. Plans now tick off step by step while Atlarix builds them.

Zed and JetBrains, smoother. The mode picker follows Atlarix when a plan is built, plans show

as the editor's own checklist, /status, /billing, /mcp and /logout work in the chat, and a

chat that can't start tells you why.

Meet Rix, the Atlarix ant. A weaver ant: the globe is any model, the compass antenna finds the

way. He heads a new terminal header (version, model, mode, plan and folder, on the left, scrolling

away with the conversation) and appears in the app, the extensions and on the site.

Back where you were. Come back to a chat after a few minutes away and one line tells you

where it stands, in the app, the terminal, Chrome and your editor. Switch it off any time

(Settings → Notifications, or /recap off in the terminal).

Privacy. For people who opted in to usage data, it now includes an approximate location

(country and city). The IP address itself isn't kept.

Fixes. You can watch the model think, and its answer appear, as it writes them in your editor

and the terminal. A reply that is just a number no longer disappears. A revoked API key now says

so, instead of "Something went wrong".

Use Atlarix inside Zed, JetBrains IDEs and VS Code. Atlarix now speaks the Agent Client

Protocol, the open standard editors use to host coding agents. Point your editor at atlarix acp

and Atlarix appears in its agent panel: the chat, the diffs and the Allow/Deny buttons are the

editor's own, and the work happens with your Atlarix account, models and settings. Every edit and

command still waits for your OK. In Zed, add Atlarix under agent_servers; in JetBrains, use AI

Assistant → Agents → Add Custom Agent; in VS Code, use the ACP Client extension. An editor

session can also hand a browser task to Chrome and get the answer back, like the terminal and the

app.

Fixes. Typing quickly in the app's message box could freeze the chat with an error; it no

longer does. Two editor windows opened back to back no longer end up sharing one conversation.

Hand a browser task to Chrome from your terminal or the app. Atlarix's Chrome extension is

now one of your sessions. From the terminal or the app, ask Atlarix to send Chrome a job, such as

"list this release on a launch site" or "draft these emails". The Atlarix icon in your browser

shows a badge; click it and the side panel starts on the task. You can steer it as it works

("shorter, less salesy") while your coding session keeps going, and Chrome sends the result

back when it is done. It works the other way too: from the browser, ask Atlarix to send

something to your terminal or app. Sending to a session that did not ask always waits for your

OK, and Atlarix still stops before any final submit.

More room on the free tier when you are new. Free Atlarix Auto now gives new accounts twice

the daily allowance for their first week, and when the full-speed share is used it says so

("Free, slower model") instead of quietly getting slower.

Pay as you go, front and centre. When the free share runs out, Atlarix offers to add credit

(from $10, no subscription) or use your own key first, and Pro after. The Chrome extension's

out-of-credit message does the same.

Atlarix can look at images, video and audio. It can open a screenshot or a chart it made,

pull frames out of a video and look at them, or transcribe a recording, in the app and in the

terminal. On Atlarix Auto it moves to a model that can see as soon as an image arrives.

Your sessions can message each other. A terminal session and an app session (or two of

either) can hand work back and forth on your computer, and the message shows as a note from

that session, never as something you typed. /handoff continues a terminal session in the app,

and "Continue in terminal" opens an app session in the terminal. Turn it off under

Settings → Notifications.

Web search on Atlarix models. With credit or Pro, Atlarix searches the web through Atlarix

itself first, which is faster and more reliable, and falls back to the free search.

The app stays light on long tasks. A long Build turn used to keep growing in memory; it now

sends only what changed, keeps command output bounded and folds older steps. The right-hand

panel no longer covers the chat.

Atlarix in Slack shows its work. One status message per task updates as it goes, from its

plan and steps to the pull request, merged or failed with the reason. Attach a screenshot and

the run gets it, tag teammates to get a private channel for the task, and ask it questions like

"why didn't my last task run?".

Also: local models from Ollama or LM Studio need no API key, temporary files land in

Atlarix's own scratch folder, atlarix --attach sends files with a scripted task, reasoning

models no longer get a temperature they ignore, and a fresh npm i -g atlarix never finds the

package without its binary.

Sub-agents work in the terminal. They failed to start before; now they run on the model

you're using, and each one shows on its own row under the prompt box: what it's doing, for how

long, and its tokens. Press ↓ and enter to watch one work, and ← → to switch between them.

Each mode follows its own instructions in every project. Outside Atlarix's own code, the

terminal app was missing the instructions behind Build, Plan, Ask, Debug and Review, and

sub-agents couldn't start. It carries them with it now.

Read the plan while you decide. A plan waiting to be built stays pinned above the prompt

box. Type what to change and watch it update, then press "Build it" or type /build.

Code is coloured. Code in answers and in edits is syntax highlighted, for whichever

languages you actually use, and inline code stands out in lavender.

Sending is instant. The first message no longer pauses for a second before it appears.

Clearer at a glance. The context meter shows tokens used out of the model's limit, long

session names are shortened, a message you send while Atlarix works shows where it read it

instead of under the reply, /compact typed mid-answer waits for the answer to finish, and

pasting the same text again expands it so you can edit it.

Also: replies draw without flicker or cut-off lines, the terminal tab shows when Atlarix is

working, Atlarix checks what your project is built with instead of assuming package.json, and

npm i -g atlarix installs two packages instead of forty.

Atlarix runs in your terminal. The same Atlarix as the app: your sessions, models, keys,

MCP servers and automations are all there. Install it with one line:

• macOS and Linux: curl -fsSL https://atlarix.dev/install | sh
• Windows: irm https://atlarix.dev/install.ps1 | iex
• Homebrew: brew install amariahak/atlarix/atlarix
• npm: npm i -g atlarix

Then type atlarix in any project. It keeps itself up to date. Sign in from your browser, or

with atlarix login --device on a machine without one.

Atlarix can work on your code while your computer is off. Ask from Slack or from

atlarix.dev/cloud. Atlarix does the work in your repository's own GitHub Actions and brings

back a pull request; nothing is merged until you approve it. It runs on paid Atlarix models,

or free on your own API key.

Relics: pages to share. Ask Atlarix for a report, a plan or a design as a relic. On a paid

Atlarix model it gets a private link you can share when you're ready; otherwise it is saved in

your project. A new Relics tab in the app lists them all.

Sign in with email. If you'd rather not use Google, GitHub or Discord, you can now create an

account with an email and password.