Atlarix

A lot shipped since v8.4. The headline change: the entire parser is different. oxc-parser and WASM tree-sitter are gone. The new stack is Universal Ctags (symbol index across 18 languages) + ast-grep (import/call/HTTP route edges via AST pattern matching) + BM25 (semantic symbol ranking) + ripgrep (text fallback). Better coverage, reliable packaged builds, and HTTP route edges as first-class structural nodes. get_blueprint is now section-scoped — focus_path required, workspace-root rejected. For a 99-file TypeScript section in a 25-section monorepo: ~6,500 provider-billed tokens, ~3s parse. Before any write, a BFS over flow edges surfaces blast radius in the approval queue. Post-turn tool-result summarisation compresses large file reads 95–98% before persisting to history. The model already saw the full output — this only shrinks downstream context cost. We published a technical paper on the context management approach with honest benchmark numbers (counterintuitive finding included): DOI: 10.5281/zenodo.20381860 v8.5 also shipped multi-terminal and workspace session history. Download: atlarix.dev

Atlarix is an AI agent workstation for software developers — runs beside your editor instead of replacing it. What's new in v8.4.0: CLI: Install atlarix from Settings and open any project with `atlarix .` from your terminal, just like `code .` Terminal: Reliable command completion with kill-on-timeout, plus an interactive terminal panel in the Activity tab Agent modes: Per-mode system prompts and per-message mode switching — each mode now has its own tool surface and identity Session memory: Edit and resend messages without duplicate user rows, with proper context replacement on continue Git: Agent commits use atlarix-agent identity with live co-author toggle in Settings Free tier works with Ollama and LM Studio — no API key needed, nothing leaves your machine. atlarix.dev

Native desktop AI coding environment — Mac + Linux. Parses your codebase into a Blueprint graph for RAG (~5K tokens vs ~100K). v8.0.1 ships: ↓ ~50% input token reduction per turn in large workspaces 🧩 Skills Marketplace — 30 community skills, open registry, Apache 2.0 🗜️ Model-triggered context compression (compress_context tool) 🔴 BYOK error banners — 401/429/404/network with fix suggestions 🔍 Explore subagent — read-only codebase exploration via delegation ✅ Streaming terminal guarantee — no more stuck loading indicators 📡 OpenRouter app identification fixed Multi-provider: any cloud API or fully local via Ollama/LM Studio. 3 tiers: Free / Pro $19 / Workforce $79. atlarix.dev

Windows support is live. If you've been waiting — it's ready. Download at atlarix.dev, and if Windows shows a SmartScreen warning just click More Info → Run Anyway. Also added 4 new AI providers: NVIDIA NIM, Cohere, Perplexity, and Cerebras. Cerebras in particular is insanely fast inference if you haven't tried it. Full platform coverage now: Mac (Apple Silicon + Intel), Linux, Windows.

Atlarix is a native desktop AI coding copilot that parses your codebase into a live architecture graph (Blueprint) using Tree-sitter. The AI queries structure before reading files — typically 5K tokens instead of 100K+. BYOK, local models via Ollama/LM Studio, free tier. What shipped in v7: The agent system is completely rebuilt. Each specialist agent (Architect, Builder, Reviewer, Debugger, Researcher) now has its own skill file that defines its capabilities and constraints. You can override any skill per workspace. A task ledger now sits outside the context window — agents always know what steps are done and what remains, even after compaction. Multi-step work no longer falls apart at context reset. Parallel execution: Research and Architect run simultaneously in Autonomous mode. Reviewer and Debugger verify after builds complete automatically. Counsel: hard design questions now run Fast, Balanced, and Thinking models in parallel then synthesize a response. Works with every provider including local Ollama. Visual git: every code change records a snapshot of the architecture diagram at that moment. You can see how your system evolved over time. Self-correction is now built in: TypeScript errors, test failures, and dev server crashes trigger automatic fix attempts with capped retries. Won the Bonus Blog Post Prize at the Devpost × AWS Nova AI Hackathon the same week we shipped. macOS + Linux. Free tier. atlarix.dev

Atlarix 6.0 is the biggest release yet. Here's what changed: Autonomous Mode — Agents now execute approved tools without interrupting you for every step. You set the scope, they work, you review the summary before anything commits. Workforce Session Scope — Connected tools (Slack, Jira, GitHub, etc.) now use per-chat workspace selections stored in memory. Start a new chat, fresh context. No more stale pre-configured defaults. Two-Stage Auto Routing — Atlarix now picks Fast / Balanced / Thinking intelligently using heuristics first, then a fast model classification pass when confidence is low. It can also suggest which specialist agent (Architect, Builder, Debugger, Reviewer) to delegate to. Linux — fully working — Packaging, OAuth deep links, Ollama/LM Studio PATH resolution, sandboxing. Linux is now a first-class platform. Thoughts Panel — See what agents are actually doing in real time via the activity stream. No more black box. UX fixes from beta feedback — Fullscreen layout, New Chat always visible, Blueprint canvas hit targets, streaming phase/tool labels, Compass usage inline with milestone warnings. Available on Mac & Linux. Windows coming soon.

v5.2.0 is out. The big one: SQLite migrations are now fully embedded in the app bundle. No more runtime path resolution, no more silent failures on first launch in packaged builds. If you hit a "no such table" error or the app felt broken on install — this fixes it. The old migration system has been replaced entirely with embedded SQL via Vite, with a bridge for existing installs so nothing gets wiped. Also shipped in the last few days: - v5.1.5: ensure* safety nets now run after migrate() instead of before, fixing fresh DB bootstrap failures - v5.1.4: auto-recovery on corrupt DBs — detects, deletes, retries once - v5.1.3: better migration path resolution for packaged apps - v5.1.1/5.1.2: MCP table bootstrapping and journal seed fixes Basically a full week of stability work under the hood. The product is in a much cleaner place now. Free to try at atlarix.dev

We just shipped the biggest update since launch — a full modal architecture replacing the old Ask/Build toggle with 5 intelligent work modes: Ask, Plan, Build, Debug, and Review. A lightweight routing agent now decides how to use your AI model based on what you're actually trying to do — so even lighter models perform significantly better. We've also built the foundation to support African-built LLMs as first-class providers, not afterthoughts. Native desktop app — Mac, & Linux.

Two big releases since we last checked in. v4.0 — Workforce The headline: AI agents that can actually authenticate with your real tools. Auth0 Token Vault is now the auth layer for all integrations — Slack, Jira, Linear, Sentry, Notion, Google Calendar, GitLab CI, Bitbucket Pipelines. Agents access these services on your behalf with scoped permissions. High-stakes actions (triggering a pipeline, posting to Slack) require step-up authentication before executing. Also shipped: Workforce tier for teams ($79/mo) — shared Blueprint, team memory, up to 20 seats. Workspace Context injects your live environment into every session: open PRs, failing pipelines, assigned tickets, Sentry errors, today's meetings. File Revert landed too — "Revert to here" on any assistant message restores your workspace file state from that point. v4.2 — Bug fixes and security Onboarding friction removed — sign-in and Skip buttons no longer hidden under window controls. Simpler, more secure sign-in flow. Legacy backend pieces removed for better stability. Pricing: Free / Pro $19/mo / Workforce $79/mo Mac + Linux. Windows coming soon. atlarix.dev

v3.9 is the biggest quality-of-life release we've shipped. Here's what's new: AI Clarifying Questions Instead of guessing your intent on complex tasks, Atlarix can now pause and ask up to 4 targeted questions before proceeding — architecture decisions, tech stack choices, destructive operations. You answer via a chip widget, answers get added to context, and the AI continues. One round only, never repetitive. Conversation Revert & Message Edit Edit any previous message and re-prompt from that point. The conversation truncates cleanly from there with a visual divider. Assistant responses that wrote files also show a "Revert to here" button — conversation revert is live now, file revert ships in v4.0. Stream Tools Two new AI tools that let the model watch live data instead of reading full dumps: stream_terminal_output — watches terminal output for a specific pattern after a command runs stream_pipeline_logs — monitors live GitHub Actions job logs Both are token-efficient and AI-initiated. Available in Ask and Build modes. GitHub Actions Panel (Pro) View workflow runs, expand jobs and steps, stream live logs, and send output directly to the AI — all without leaving Atlarix. Lives in the left panel alongside DB Studio. Requires GitHub sign-in. Resizable Left Panel Drag the right edge of the left panel to resize. Min 200px, max 60% of window. Your preferred width is remembered across sessions. Chat Tab When Panels Cover the Screen When both panels are open and the main chat area is hidden, a Chat tab appears in the right panel — full session, same messages, same input. Model selection uses a compact inline search instead of the full sidebar. Panel Keyboard Shortcuts (Pro) Cmd+I / Ctrl+I — toggle left panel at 30% width Cmd+O / Ctrl+O — toggle right panel at 30% width Removed Prompt suggestion chips. Atlarix is a developer tool — you already know what you want to ask.

Two things this update: DB connections got simpler. Paste a connection string and Atlarix auto-detects the engine (PostgreSQL, MySQL, MariaDB, SQLite) and reads the database name automatically. Hit Parse, hit Test Connection, you're done. No more manual configuration. Onboarding is now a real walkthrough. Instead of tooltips, v3.8 walks you through actual app screens — workspace setup, your live Blueprint, Ask vs Build mode, and adding your AI provider. Your real project is the demo from step one. Also: GitHub OAuth is now available alongside Google sign-in. Mac + Linux. Free tier available. Pro at $19/month.