The State of Atlarix: What We Cut, Kept, and Sharpened
If you read our earliest posts, you'll see ideas that no longer ship: a Blueprint structural-graph index, model tiers, a separate live-preview pane. That's not drift — it's the plan. Atlarix has spent the last stretch refining, and the rule the whole time has been simple: simplify the surface, never the capability. The product you install today is smaller in moving parts and larger in what it can actually do. Here's where Atlarix stands now, and the honest story of what we cut to get here.
| Earlier Atlarix | Atlarix today | |
|---|---|---|
| Codebase context | Blueprint structural-graph RAG + an in-memory index built on open | Bundled ripgrep (grep + glob) — no index, nothing built when you open a workspace |
| Models | fast / balanced / thinking tiers | No tiers — a live models.dev Core catalog, any provider via BYOK, or fully local |
| Core billing | bundled into a tier | Pay-as-you-go: a prepaid credit wallet (from $10) on every plan, metered as you go |
| Running your app | a separate Preview tab | It runs in the in-app Browser tab — one surface, agent-drivable |
| Runtime errors | a preview error panel | The Errors tab, now with full stack traces |
| Command safety | hardcoded 'safe' and 'mutating' command lists | A per-OS write sandbox + auto-run + ask-once escalation on denial |
| Tool surface | ~38 agent tools | ~22 focused tools |
| Editor stance | drifting toward an IDE | Not an IDE — a workstation beside the editor you already use |
The refinement principle
Every removal below follows one rule: cut surface, keep capability. A feature that duplicated something the model already does well, or that added machinery competing with the model's own judgment, was a candidate for the chopping block — not to make Atlarix smaller for its own sake, but because a leaner harness is a more reliable one. Where the reference harnesses (Claude Code, opencode, Codex, Cursor) had already converged on a simpler primitive, we stopped inventing our own and matched what works. The scoreboard: the agent's tool count went from ~38 to ~22, and whole subsystems came out — thousands of lines deleted — while the things you actually do with Atlarix got broader, not narrower.
Retrieval: Blueprint retired, ripgrep won
We shipped Blueprint — a section-scoped structural graph that gave the model a compact map of a codebase — and even published a benchmark on it. The research still stands; it measured a real effect. But as a product decision, we retired the Blueprint tool and tore down the in-memory index behind it (the BM25 store, the file watcher, the symbol tagger). In their place is the primitive Claude Code and opencode already trusted: bundled ripgrep. `grep` searches contents, `glob` finds files, and the model reasons over the ranked results. Nothing is built when you open a workspace, nothing drifts as files change, nothing is uploaded, and retrieval stays fast and flat-memory on a repo of any size. The counterintuitive lesson of our own research — that more structure didn't reliably mean better answers — is exactly why the simpler tool won.
One workstation, fewer surfaces
The most recent refinement (v14.15.0) removed the Preview tab. For a while Atlarix had two places a web app could live — a Preview pane and the built-in Browser — and 99% of the time they were the same thing: a URL the agent could drive. So we collapsed them. Your dev server now opens in a Browser tab like any other page; the agent navigates, clicks, fills, and screenshots it, or you do. Runtime errors — console errors, uncaught exceptions, failed requests — surface in the Errors tab with real stack traces, captured straight from the running page so the agent can see and fix what actually broke. One surface to learn, and it does more than the two it replaced.
Models: no tiers, pay-as-you-go Core
The old fast/balanced/thinking tiers are gone. Atlarix Core is now a live catalog pulled from models.dev — DeepSeek, Qwen, Kimi, MiniMax and friends stay current without an app update — and you pick the exact model in the picker. Core is pay-as-you-go on every plan: top up a prepaid credit wallet (from $10) and it's metered as you go, at real cost. Bring your own key for any provider, or run fully local with Ollama and LM Studio, free on any plan. No tier decides which models you're allowed to think with.
Safety that stays out of the way
We deleted the hardcoded taxonomies that used to guess which commands were 'safe' and which 'mutating.' Enforcement is now structural: commands run inside a per-OS write sandbox — macOS Seatbelt, Linux Landlock, Windows AppContainer — confined to your workspace and its caches, with the network left on and reads left broad. Ordinary commands just run; only genuinely dangerous ones pause to ask. And when a real toolchain needs to write somewhere new, Atlarix asks once, grants that exact path, and re-runs the command still sandboxed — so a first-run `yarn install` no longer dies on a missing cache dir. File writes still wait for your hunk-level diff approval, and read-only modes can't write at all.
The agent loop, hardened
Underneath the surface changes, the loop itself got tougher. Writer sub-agents draft edits in throwaway git worktrees and hand their diffs back for approval — parallel throughput without surrendering review. Long commands move to the background after ~15 seconds and deliver their result when they finish; you can even put a live countdown on a background job and watch it tick down. After each round of edits, Atlarix runs your project's own checks (tsc, eslint, ruff, pytest…) and feeds failures back to itself, so a task can't be marked done on top of broken code. The loop ends when the agent is actually stuck, not when a step counter runs out. And because open-weight models are first-class here, the harness prefers native function-calling and falls back to a text/XML tool protocol when a provider needs it — the reason local and open models actually perform in Atlarix where they stall elsewhere.
What didn't change
The thesis is the same one we started with: the open-weight frontier, local-first, with your approval before every write, beside whatever editor you already use. Atlarix is deliberately not an IDE — it doesn't run a language server or reimplement a Problems panel; your editor already does that. What refinement changed is everything around that thesis: fewer tabs, fewer tools, no index to build, no tiers to navigate — and a browser, a real terminal, MCP servers, automations, voice input, and concurrent sessions that all got sharper in the process.
The honest version of a changelog is that some early bets don't survive contact with real use — and saying so is the point. Blueprint, tiers, and the Preview tab were good ideas that a simpler design outperformed. What's left is a harness that does more with less: open-weight and local models, no-index retrieval, a per-OS sandbox, a browser-native run-and-debug loop, and approval before anything touches disk. That's the state of Atlarix today — and it's the most capable it has ever been.