Engineering Blog

Run Claude as your coding agent's model however your company buys it: your own Anthropic API key, Amazon Bedrock's Anthropic endpoint, or a gateway that speaks Anthropic's Messages API. How to connect each in Atlarix, what to type where, and what works the same on every route.

Atlarix marks its system prompt for the model's prompt cache, and inside one message the cache worked. Across messages it never did: each new message in Build mode wrote about 19,000 tokens to the cache instead of reading them. One block of text was in the wrong place. Here is how we found it, the numbers before and after, and the one-line rule it comes down to.

Atlarix now runs inside your editor. One extension for VS Code, Cursor, Kiro and Windsurf carries the whole agent: every mode, any model or your own API key, edits in the editor's own diff, plans that tick off as they are built, and your Chrome. Zed and JetBrains IDEs get the same agent over the Agent Client Protocol.

Bring your own key from OpenAI, Anthropic, Google, DeepSeek, OpenRouter or another provider, and run a full coding agent in VS Code, Cursor, Kiro or Windsurf on it. Step by step with the Atlarix extension, what your key costs you, and where it is kept.

Zed and JetBrains IDEs speak the Agent Client Protocol, so any agent that implements it runs in their own chat panel with no plugin. Here is how to add Atlarix to both — the exact settings, what shows up in each editor, and how sign-in works.

Atlarix 15.0.1 is the first update to the Atlarix CLI, and most of it came from using it on a real pull request. Sub-agents now run in the terminal and sit under the prompt box, code is syntax highlighted for the languages you actually read, the plan stays pinned while you change it, and sending is instant.

No API key, no credit, no model to choose. Atlarix Auto routes every step of the work — exploring, building, planning, reading a screenshot — to a model that fits it. Free in the desktop app on accounts with no credit, private on Pro. Here is exactly how it works, including the trade that makes the free part free.

Round-Trip Engineering, a live SQLite Blueprint graph, 57 tools, a $79 Team tier, Windows coming soon — none of it describes the product. Some was true once; some we published in error. Every stale claim, corrected and dated, in one place.

Anthropic prices pull-request review at $15-25 per review, billed separately and limited to Team and Enterprise; browser control needs a direct Anthropic plan. Atlarix includes both in one subscription. A comparison sourced entirely to Anthropic's docs.

We drafted this post around Cursor indexing your code into embeddings. Then we checked their live documentation and that claim was gone — replaced by Instant Grep. Here is the honest comparison that survived fact-checking.

GitHub documents a semantic code search index, and for repositories hosted elsewhere states plainly that it uploads your data to make it searchable. Atlarix builds no index at all. That difference decides most of the rest, including who pays for what.

Codex is Apache-2.0 and runs local models through Ollama and LM Studio, so claiming local support as an Atlarix advantage would be false. The real differences are what surrounds the agent and how review is metered.

opencode documents permissive defaults where most permissions allow automatically; Atlarix queues every write for approval. Two coherent bets about what interruption is worth. Plus review in your own runner versus a reviewer that cannot execute anything.

Not a policy summary — the mechanics. What Atlarix never stores, what genuinely does leave your machine and why it has to, how approval and the OS sandbox work, and the part of the Chrome extension we would rather you heard from us.

Chunk, embed, store, invalidate, retrieve by similarity — the reference architecture. Atlarix does none of it. The case for lexical search on code, including where it genuinely loses, and why we removed an index we had already shipped.

Ollama or LM Studio, a tool-capable model, and about ten minutes. How to connect one to Atlarix, why retrieval matters more on a constrained context window, and the two places local still disappoints — including two of our own surfaces.

FTS5, BM25, ctags, a reranker and a file-watcher — the recommended architecture, removed in one release. The watcher was holding one file descriptor per watched file and reaching about 12,365 on a large repo, until the app could no longer spawn a process.

Seatbelt, Landlock and — after a rewrite — Windows Low Integrity Level. The first Windows attempt used AppContainer, compiled fine, and broke bash, git and PowerShell the first time it touched real hardware. Why, and what we took from it.

We won a Terminal-Bench comparison by three tasks and have spent more effort explaining why that proves nothing. Also: the denominator on our own benchmark page was wrong, roughly a quarter of trials timed out, and no Atlarix SWE-bench number exists.

Hitting Stop used to bill you for output the model never produced — the whole budget it was allowed, not the two seconds it actually ran. The same request measured 45c completed and 58c cancelled. Here is what was wrong, what it now does, and why we think a meter you cannot audit is worse than an expensive one.

We ran a controlled benchmark comparing Blueprint-assisted vs unassisted agentic exploration on a production multi-repository workspace. Here's what we found — including a counterintuitive result about total context usage.

Everyone reaches for embeddings and a vector database. Atlarix doesn't — no embedding model, no vector index, nothing uploaded. Here's how bundled ripgrep and the model's own reasoning give it real, whole-codebase context, and why that's the better call for code.

Benchmarks, demos, and feature lists don't tell you if a tool works on your codebase. Here's the checklist we use — and that you can use — to evaluate any coding agent on what actually matters: model access, retrieval, control, and data path.

Run the full coding workflow on your own hardware with Ollama or LM Studio. Here's why local models actually perform in a desktop workstation, where they often don't elsewhere.

Most AI coding tools ship as editor forks or plugins. Atlarix is a standalone desktop workstation that sits beside any editor. Here's the capability case — local models, the approval queue and sandbox, and a browser/terminal around the agent.