Atlarix

Since the last update here (the Network tab, v14.31), Atlarix has shipped through the 14.3x line. This is the through-line — full notes for every version are on the changelog. It checks its own work now. After building something, Atlarix hands a separate worker the requirement and the running URL — not the diff — and has it exercise the app independently, with a real terminal and browser but no ability to change a file. It reports what it saw with evidence: a status code, a console line, a log excerpt. Everything it claims, Atlarix reproduces before acting on it. Debug mode also got the full build loop, so it runs your project's own checks and won't call itself done while they report errors. Long sessions stop rotting. Compaction was three implementations with three thresholds, two of which threw their work away every turn. It's now one system: it summarises before deleting anything, scales to the model you're on instead of a fixed token count, and keeps the last thing you said intact. A new session_search tool reads back through the full transcript on disk — a path you gave once, an error string, a decision twenty turns ago — so the agent looks it up instead of asking you to repeat yourself. A dropped connection no longer ends a run. A provider hiccup used to skip the retry logic and kill the whole turn half-finished. Those now hit the retry that was always meant to catch them; genuine faults still stop the run. Work mode belongs to the conversation, not the window. Plan in one session and Build in another used to be impossible. Each session remembers its own mode now. Plus: a Diagnostics screen, deep thinking that reaches the provider, honest web search, readable scanned PDFs, and Core 1 on GLM 5.3. Full notes: atlarix.dev/changelog

New product, same account: Atlarix Reviewer reads every pull request on the repositories you pick and posts its findings as a review — file, line, what's wrong, and the fix — as review comments, not a score. Two decisions carry it, and one of them is a refusal. It never executes your code. It reasons about the diff and never builds, runs, or tests it — there is no runner to compromise and no build step to poison. The cost is real: anything that only shows up at runtime, it structurally cannot catch. The trade is deliberate. A reviewer that quietly reports a pull request clean because its test harness failed to start is worse than one that says plainly it only read the code — so this one only ever claims to have read the code, and says so when a review didn't finish instead of calling the PR clean. The second decision is about crying wolf. Every candidate finding is checked a second time before it posts. The ones that don't survive that check are counted, not shown. A reviewer that flags everything suspicious gets muted, and a muted reviewer reviews nothing — so the second pass exists to protect the one thing the product is for. It's also the roughest edge right now, and the part I most want stress-tested: false-positive tuning is exactly where an early reviewer is weakest. What it asks GitHub for is narrow: read on contents, write on pull-request reviews. It never pushes, never merges, never runs tests. $19 per developer per month, unlimited repositories — per developer, never per repository, because charging per repo punishes the teams with the most PRs and the most to gain from reviewing them. Installs on your own repos in a couple clicks. Throw a PR at it and tell me where it over-flags. → atlarix.dev/reviewer

Two changes this release, and the second is what made the first worth building. Nothing the browser sees can end a turn any more. A failed request or console error in a page used to be able to stop the agent and nag it to fix what it found — but Atlarix can't tell whose error it is. A frontend on :5173 calling its own API on :8080 is one app across two origins; an embedded payment script is someone else's code until the day it's yours. Every rule we could write was wrong in one direction. In practice it went wrong the expensive way: the agent opened YouTube just to read it, hit Google's signed-out 401, and was told to go fix a third-party login handshake. It refused — correctly — then burned turns arguing with something that couldn't hear it. So now the agent is told, not blocked: one line per turn, counts and where to look, and it reads the detail itself only if it matters. Settings → Browser decides what it's told — everything, only failures, or nothing. A real Network tab. Every request each browser tab made — method, status, type, duration, size — with headers and response bodies when you expand a row. One card per tab, so what YouTube did and what GitHub did can never be read as the same thing. Send to Atlarix on a row, a tab, or everything shown — added behind a separator so it never eats what you were typing. The Console tab groups by tab too, and navigating no longer clears it: comparing two page loads is often the whole point. Fixes: "Restoring your session…" no longer hangs forever when you started signed out. A Linux crash from desktop notifications is gone. Reconnect retries now jitter instead of firing in lockstep, and sub-agents start staggered so they share the prompt cache instead of each paying to write it. BYOK, run local, or Core credits. → atlarix.dev

→ Atlarix is live on the Microsoft Store. One-click install, no "unknown publisher" wall — Store installs keep your existing workspaces, chats and credentials, and update through the Store instead of the in-app updater. → Store submission is now fully automated — every release tag ships to the Store on its own, so updates land faster and more reliably than the old manual path. → Windows sandbox hardened — toolchain caches outside your user profile (npm, pnpm, yarn, Go, Deno) are writable again, and LocalLow is resolved properly so redirected-AppData machines stop hitting "cannot create a temp file." Bring your own key, run local, or use Core credits. → apps.microsoft.com/detail/9p3rr11dvzpw

This release is mostly about Atlarix no longer doing things to you that it shouldn't. Signing out no longer logs you out — or wipes your setup. Two quiet bugs got fixed for good: when the OS keystore was briefly busy, Atlarix could read that as "no data" and overwrite the real thing — costing you your session or your saved API keys. And signing out used to clear every workspace, chat, and automation under a "fresh start" that nobody asked for. Both are gone. Unreadable never means deleted now, and wanting your data cleared is an unchecked box, not the default. Nothing you didn't ask to remove gets removed. On Windows, search actually works. Three path-handling bugs meant grep could silently find nothing, folder listings returned your whole project, and file globs errored out — all only on Windows, all now fixed. If you're on Windows, the assistant now sees your code the way it should. The AI understands your automations and your plan. Scheduled tasks, triggers, and runners are now visible to the agent, so it can actually reason about them — and it always tells you whether an automation will wake the AI (and cost a turn) or just run quietly on its own. It also knows your plan's limits now, so "you can't add another" comes with a clear reason instead of looking like something broke. Automations got a real schedule editor, too — no more hand-writing cron. Lighter and more honest under the hood. The context meter now counts what the model actually holds instead of guessing, skills load as a lightweight index rather than dumping everything into every message, and a big pile of dead code came out. Faster on large repos, and clearer about what it's doing. Pasted-text sessions get real names now, @ leads with your files, and the agent's terminal output stays in chat where it belongs.

Two releases since 14.15, and they share a theme: no network, a background tab, a provider returning half a list — none of those should be your problem. The browser agent grew up. It can hold several tabs open at once, so opening a second page no longer destroys the first one's state — a signed-in session, a scroll position, a half-filled form. It can find an element by describing it ("login button", "search box") instead of reading the whole page first. And — embarrassing in hindsight — it can now press Enter. It couldn't before: no submitting a search box, no Escape to close a modal, no Tab between fields. You can also watch it work now: a cursor glides to each target and stays on screen, with a label naming what it just did. The app opens offline, in the time it takes to draw a window. It used to hang for minutes on a captive-portal or hotel wifi, which reads as "the app won't open". And an offline Pro user is no longer quietly dropped to Free for the session. Model lists can't shrink. A public model catalog is now the floor under every provider, so a truncated response from a provider can no longer make models vanish from your picker. Dictation is materially more accurate — bigger on-device model, the mic's call-tuned processing turned off, and your spoken language passed in rather than guessed at. 14.16.1 on top of that: you stay signed in across restarts (a 14.16.0 regression, my fault), a provider's models appear the moment you save its API key, sign in with Discord alongside Google and GitHub, and a new Settings → Community with the Discord invite and a link back here.

→ One surface for your running app. The separate Preview tab is gone — your dev server now opens in a regular Browser tab the agent can navigate, click, fill, and screenshot. One place to learn instead of two that did the same thing. → Runtime errors with real stack traces. The Errors tab now captures uncaught exceptions and unhandled rejections with their stack, straight from the running page — so the agent sees and fixes what actually broke, not just that something did. → Clickable URLs. Links in chat and the terminal — including bare localhost:PORT — open in a Browser tab (⌘/Ctrl-click for your system browser). → Sandbox that doesn't fight your toolchain. When a real tool needs to write somewhere new, Atlarix asks once, grants that exact path, and re-runs the command still sandboxed — so a first-run yarn/pnpm install no longer dies on a missing cache dir. No enumerated cache list to keep complete. → Live countdown on background jobs. Long-running commands moved to the background now show a wake timer you (and the agent) can see and adjust — set it, watch it tick, re-check when it fires. → Quieter + cleaner. New keyboard shortcuts (Cmd+T/W/J/I/O), device-spec awareness so the agent sizes parallel work to your machine, and a code-block render fix. Download for Mac, Linux & Windows → atlarix.dev (Windows unsigned; signing coming.)

→ Web search on Atlarix Core. Core models can now search the web themselves, pay-per-use through your credit wallet — no separate setup, no always-on plugin. (Citation strip lands in the next patch; results already work.) → Work across repos, not just one folder. The workspace is now a smart default, not a wall — the agent can read, grep, and edit files outside the current workspace (edits still go through the approval gate). No more shell-edit workarounds for dependency or cross-repo work. → Auto-update actually works now — all three platforms. macOS, Linux AppImage, and Windows auto-update were rebuilt on the correct native path for each OS. (One-time catch: installs on 14.13.2 or older need a single manual update to 14.13.3+, then it self-heals forever.) → Real model names on Atlarix Core. The Core provider page shows live names for Kimi , Qwen, DeepSeek, MiniMax , instead of "Core N" placeholders. Core badge now reads "Setup Complete" (no key needed). → Cleaner chat + honest tool cards. Injection-defense envelopes no longer leak into the UI. Markdown tables and inline code render correctly. Negative wallet balances explain themselves. → Reliability batch. Permission writes can't wedge a turn on a full disk; sandbox respects custom pnpm/cargo/etc. stores (fixes stalled installs); dead Core metering error removed; leaner agent prompting. → Faster model list. The 167-provider catalog no longer rebuilds on every window focus.

-> Pricing reframe — Free + Pro + PAYG credits. No tier bundles AI. Atlarix Core is now pay-as-you-go prepaid credit on every tier — top up a wallet, spend it across any model. Free tier (BYOK/local), Pro ($19), and PAYG credits that work on all of them. Team tier removed. -> Real model pricing in the picker. Every provider's live per-token cost (via models.dev + OpenRouter's own pricing) now shows as a real $/M badge, so you see what a model actually costs before you run it. -> edit_file overhaul. Atomic multi-edit support plus Cursor-style merge with `// ... existing code ...` markers — deterministic stitching that refuses rather than corrupts. -> Co-author, not author. Commits are authored by your own git identity; Atlarix is credited only via a `Co-authored-by:` trailer, and it's fully toggleable. -> Harness reliability batch. Smarter wait/stall handling, an always-on live token counter, graceful image-model fallback, and paste/first-run fixes — tuned for driving weaker open-weight models without stalls. -> Security hardening. Plaintext-keychain hardening (chmod 600 + re-encrypt-on-recovery), privileged-IPC gating, and workspace-boundary read/write confinement.

New pricing model. No more capped free token pool. Run Atlarix free with your own model — DeepSeek, Qwen, or local via Ollama — or top up pay-as-you-go credits and pay only for what you use. Pro unlocks unlimited workspaces and the full tooling set. Voice-to-text. Talk to your codebase — free and local via Whisper. Performance. Message virtualization and a faster streaming pipeline keep the UI responsive on large repos and long sessions. Also in this release: live token counters, attachment dedup, and a range of stability fixes.