Atlarix: The Open-Weight Frontier Harness for AI Coding
Atlarix: The Open-Weight Frontier Harness
Most AI coding harnesses treat open-weight models as an afterthought — a model picker dropdown bolted onto a UX designed around Claude or GPT. Atlarix is built the other way around. It's a desktop coding harness designed from the ground up for open-weight frontier models: DeepSeek, Qwen (Alibaba), Kimi (Moonshot), and MiniMax.
Why it's different: Blueprint
The core of Atlarix is Blueprint, a structural + lexical code-retrieval engine — no embeddings, no vector DB, no reranker. It builds a real dependency graph of your codebase and hands the model structurally-grounded context instead of a nearest-neighbor guess.
This matters most on exactly the models Atlarix targets. Open-weight models generally have less forgiving context discipline than the top closed models, so retrieval quality is often the difference between a model that's genuinely useful on a real codebase and one that quietly hallucinates around the gaps.
Five modes, matched to the task
Atlarix runs on a five-mode workflow — Ask, Plan, Build, Debug, and Review — so the model's autonomy always matches what you're actually trying to do, instead of one flat "agent mode" that's either too cautious or too reckless for the moment.
Approval-first, not approval-optional
"AI that asks before it acts" isn't a tagline, it's the execution model. Every file write, every terminal command, every MCP tool call routes through a real permission gate — not a checkbox everyone disables on day one. MCP tool descriptions are treated as untrusted input, OAuth uses PKCE, and there's a live consent gate on tool calls, not just a static allowlist.
Automations, safely bounded
Cron, GitHub-poll, and MCP-poll triggers let Atlarix run unattended — dangerous operations are always denied in that mode, and every run has a hard execution budget and concurrency limit.
Lean by design
367MB installed. Electron doesn't have to mean bloated.
Built solo, out of Nairobi
Atlarix is built by one engineer under NorahLabs — not venture-backed, not optimizing to be the biggest. The goal is to be the most honest, best-built option for developers who've already chosen open-weight models on cost, sovereignty, or capability grounds, and want a harness that respects that choice instead of quietly steering them back to closed models.