The workstation built for the open-weight frontier.
DeepSeek, Qwen, Kimi, MiniMax — Atlarix Core ships the open-weight frontier labs, native tool-calling and fast codebase search make any model perform, every file edit still waits for your approval, and commands run sandboxed. Any model BYOK. Local-first. Built in Nairobi.
macOS · Linux · Windows · No mandatory API key
Built for the open-weight frontier
The frontier isn't closed anymore. DeepSeek, Qwen, Kimi, and MiniMax ship frontier-class models at open-weight prices — and Atlarix is the harness built to make them perform. Fast lexical search over your codebase — bundled ripgrep, no index — plus native tool-calling give any model sharp, grounded context, so an open-weight model in Atlarix works like it knows your project.
You choose how you connect:
Atlarix Core — the open-weight frontier
Four managed models from the frontier open-weight labs — DeepSeek, Qwen, Kimi, MiniMax — pay-as-you-go, metered per use, with zero API-key setup.
Your own accounts
Bring the AI providers you already pay for — OpenAI, Anthropic, Gemini, OpenRouter, or any OpenAI-compatible endpoint.
Ollama / LM Studio
Run open-weight models on your own machine for zero external upload — free, offline, air-gapped when policy demands it.
Mix and match
Choose per project, workspace, or conversation — stay flexible as needs change.
If a vendor changes terms, doubles prices, or goes down — you switch in one click. No retraining. No migration panic. Your AI strategy stays yours.
Retrieval
Why open-weight models punch above their weight here
A model is only as good as what it can see of your code. Atlarix gives any model fast, purely-lexical search over your codebase — locally, with no index and no embeddings — so an open-weight model in Atlarix works like it already knows your project.
Search, not similarity
`grep` and `glob` run on bundled ripgrep over your workspace — exact matches with real file:line grounding, honoring your .gitignore. The model gets the code that's actually there, not a nearest-neighbor guess.
No index, no watcher
There's nothing to build when you open a workspace and nothing running in the background between searches. Memory stays flat and startup is instant at any repo size — nothing is scanned until the agent searches.
The model does the reasoning
Like Claude Code and opencode, the agent reasons over ranked search results rather than a vector index or a structural map. Tool calls and results ride as provider-native blocks, so open-weight and local models use them well.
No embeddings. All local.
No embedding model, no vector database, nothing uploaded — retrieval stays purely lexical. It runs on your machine, so it scales to any repo at constant, low memory.
The payoff: no index to build or drift, retrieval scales to any repo size, and the frontier open-weight labs — DeepSeek, Qwen, Kimi, MiniMax — navigate your codebase from sharp, grounded search results.
In our first Terminal-Bench 2.0 test, Atlarix held its own against a leading agent harness on the same open model — preliminary, fully disclosed.
How it works
Connect your workspace
Open any project folder. Atlarix reads your codebase structure without uploading anything.
Ask the AI to do something
Write a task in plain language. The AI analyzes your code and builds a plan.
You stay in control
Every file edit shows a diff to approve or reject, right in the chat. Commands run inside a per-OS sandbox that can only write to your workspace — and genuinely dangerous ones still ask first.
See Atlarix in action
Watch the open-weight frontier at work.
Feature walkthroughs
Built-in Browser & Web Access
A full browser inside your workspace. Load docs, test live endpoints, and research without leaving the app. The agent can read what's on screen.
Agent-Assisted Development
Watch the agent plan, write, and verify in real time. Every file edit shows a diff you approve or reject by hunk, right in the chat — while commands run inside a per-OS sandbox confined to your workspace.
Sub-agents & Activity
Spawn scoped sub-agents to explore, plan, write code in an isolated worktree, or search the web — up to five at once — while your main session keeps running. Each one streams its thinking and tool calls into its own card, inline in the chat.
Real Terminal & Live Preview
Full interactive shells with tab completion, colours, and persistent scrollback across restarts. Preview your app live alongside the agent. Build, run, and see results without switching context.
Any model, including ones you already have
Atlarix Core ships DeepSeek, Qwen, Kimi, and MiniMax built in. Or bring your Anthropic, OpenAI, or Gemini key — or run fully offline with Ollama and LM Studio. No new subscriptions required.
Your edits wait for you; commands stay sandboxed
Every file edit shows a diff you approve or reject by hunk, inline in the chat. Commands run inside a per-OS write-confining sandbox — dangerous ones still ask, and allow-once / always-allow decisions are scoped to the workspace.
Automations that run without you
Schedule the agent on a cron, a GitHub event, or an MCP trigger — it runs unattended and reports back. Two automations per workspace on Free, ten on Pro.
Talk to it — for free
Built-in voice-to-text runs on-device with local Whisper — private and free, no key. Prefer cloud accuracy? Bring your own OpenAI, Groq, or Gemini key; it never touches your Core credits.
Open-weight economics, local-first by design
Most AI coding tools assume you have a $20/month plan, a mandatory API bill, and a fast connection.
Atlarix is built on open-weight economics instead: the Core models — DeepSeek, Qwen, Kimi, MiniMax — cost a fraction of closed APIs, and Ollama and LM Studio run entirely free on your own machine.
No cloud dependency. No recurring API bill if you don't want one. Your code never leaves your machine unless you choose it to.
Works on a slow connection. Works air-gapped. Works anywhere.
Simple pricing. No surprises.
| Free | Pro — $19/mo | |
|---|---|---|
| Price | $0 | $19/month |
| Workspaces | 1 | Unlimited |
| Atlarix Core (all models) | Pay-as-you-go credits | Pay-as-you-go credits |
| BYOK / Local Models | ✓ | ✓ |
| Approval Queue | ✓ | ✓ |
| MCP servers | 1 / workspace (manual) | 10 / workspace |
| Automations | 2 / workspace | 10 / workspace |
| Skills | 3 / workspace | Unlimited + marketplace |
| MCP & Skills Marketplace | — | ✓ |
| One-click OAuth connectors (Slack, Jira, Linear, Notion…) | — | ✓ |
| Sub-agent fleets (parallelism) | Limited | Full |
| Databases & CI/CD (git, GitHub/GitLab, psql…) | ✓ via terminal | ✓ via terminal |
Free
Price$0
Workspaces1
Atlarix Core (all models)Pay-as-you-go credits
BYOK / Local Models✓
Approval Queue✓
MCP servers1 / workspace (manual)
Automations2 / workspace
Skills3 / workspace
MCP & Skills Marketplace—
One-click OAuth connectors (Slack, Jira, Linear, Notion…)—
Sub-agent fleets (parallelism)Limited
Databases & CI/CD (git, GitHub/GitLab, psql…)✓ via terminal
Pro — $19/mo
Price$19/month
WorkspacesUnlimited
Atlarix Core (all models)Pay-as-you-go credits
BYOK / Local Models✓
Approval Queue✓
MCP servers10 / workspace
Automations10 / workspace
SkillsUnlimited + marketplace
MCP & Skills Marketplace✓
One-click OAuth connectors (Slack, Jira, Linear, Notion…)✓
Sub-agent fleets (parallelism)Full
Databases & CI/CD (git, GitHub/GitLab, psql…)✓ via terminal
Atlarix Core is pay-as-you-go
Atlarix is a harness, not an AI provider — so no plan bundles AI. Hosted models run on prepaid credit you top up (from $10), metered as you go. The plan buys the harness; credits buy the tokens. Prefer your own models? BYOK and local models are free on every plan.
What People Say
Teams and builders using Atlarix to understand and ship code faster.
Atlarix turned a codebase I barely knew into something I could actually see. Instead of reading file after file to work out how everything connects, I got a clean visual map of the whole thing: the structure, the dependencies, how each piece talks to the next. It is the fastest I have gone from 'lost in a repo' to 'I get how this works.' For anyone who has to understand code they did not write, it is a genuinely different way of working.
Naga, Co-founder, AgentSonar
Atlarix helped us develop our work systems in our organisation. A really good program.
Victor Njeru, CEO, Nilinki
“Atlarix turned a codebase I barely knew into something I could actually see. Instead of reading file after file to work out how everything connects, I got a clean visual map of the whole thing: the structure, the dependencies, how each piece talks to the next. It is the fastest I have gone from 'lost in a repo' to 'I get how this works.' For anyone who has to understand code they did not write, it is a genuinely different way of working.”
Naga
Co-founder, AgentSonar
Frequently Asked Questions
Does my code get sent anywhere?
Only if you use a cloud AI provider and choose to. With Ollama or LM Studio, nothing leaves your machine. Ever.
Does Atlarix save my chat history?
In a workspace, yes — on your machine. Messages are stored as sessions in the sidebar so you can resume where you left off. You need an open workspace to chat; there is no separate global chat. Nothing is sent to Atlarix servers unless you use a cloud model. The right-panel terminal keeps command output until you start a new chat or close that shell. MCP servers, skills, and project memory under `.atlarix/` are scoped to each workspace.
What AI models does Atlarix support?
We fully support openweight models, specifically the ones from DeepSeek, Qwen, Kimi, and MiniMax. However we allow users to bring their own models as well
What's Atlarix Core?
Atlarix Core is our included model catalog — managed open-weight models from the frontier open-weight labs: DeepSeek, Qwen, Kimi, and MiniMax. Core is pay-as-you-go on every plan: top up a prepaid credit balance (from $10) and it's metered as you go — you pay only for what you run, and the app shows the exact cost of every call. Prefer your own models? BYOK and local models (Ollama, LM Studio) are free on any plan.
Is the approval queue on every plan?
Yes. The approval queue is on every plan including free. It's not a paid feature. It's how Atlarix works.
Stop hoping the AI didn't change something important.
See exactly what Atlarix does before it does it.
Ready to Get Started?
Download Atlarix for your platform and start building smarter today.
These artifacts are the same Electron builds developers run locally: signed/notarized macOS bundles, Linux packages, and Windows installers as they roll through the public release repo. Verify checksums on GitHub before you deploy across a team.
macOS
Choose architecture
Notarized and code-signed — macOS trusts it out of the box.
Windows
Setup.exe
View on GitHubWindows Build — Uncertified
Atlarix for Windows has NOT been code-signed or certified yet. Windows Defender will show a "Windows protected your PC" screen before install — this is expected. To proceed, click "More info" then "Run anyway".
By downloading this build you confirm you understand it is currently unsigned and are choosing to proceed at your own discretion. Full code signing is coming soon.
If you need a certified build today, macOS is fully notarized and signed.
Unsigned build — Windows Defender SmartScreen will warn before install; this is expected.
Linux
Choose format
Early Access Notice
Atlarix is actively evolving.
While we continue improving workspace handling, we strongly recommend using Atlarix on duplicate or non-critical project folders. This helps ensure your primary codebases remain fully safe as we ship rapid updates.
SHA256 checksums available for download integrity verification