Bring your own model. Shape it for the work you actually need to do.
The current Windows preview is Atraium 0.2.3 Preview 1.
| Download | Purpose |
|---|---|
| Windows x64 installer (.exe) | Installs Atraium for the current Windows user. |
| Portable Windows x64 archive (.zip) | Extract and run Atraium without an installer. |
| SHA-256 checksums | Verify the installer or portable archive before running it. |
| All preview releases | Browse release history, notes, and future downloads. |
Read Getting started before connecting a provider. For the details of this version, see the 0.2.3 Preview 1 release notes.
Important: Atraium uses the provider accounts and credentials that you configure. You are responsible for all resulting provider and infrastructure charges. Read Token usage and provider billing and set provider-side budgets or limits before using paid endpoints.
Welcome. This repository is the public home for trying Atraium before any broader source release.
It intentionally contains no Atraium application source code. Instead, it provides:
- installer and portable-archive downloads through GitHub Releases;
- getting-started, privacy, security, billing, and support guidance;
- versioned release notes and known preview limitations;
- GitHub Issues for reproducible bugs and product feedback.
In other words, this is a binary-preview and documentation repository, not the Atraium source repository. You can download, install, and evaluate the preview here. Any future public source release and its licence will be announced separately.
Atraium is a local-first, bring-your-own-model workspace for people who want more from AI than a blank chat box.
You configure the providers and capabilities you want, create rooms with their own assistant identities and instructions, and give those rooms practical tools for research, files, software, documents, images, and desktop workflows.
Atraium is available as an early external Windows preview.
| Target | Preview status | Guidance |
|---|---|---|
| Windows desktop | Most stable and recommended | Download the installer or portable archive above. |
| Android | Expected to work, but less regularly validated | Not included unless an Android asset is explicitly attached to a release. |
| Web | Future work; not currently tested or supported | Do not treat the web host as an available preview target. |
Depending on your model, configuration, platform, and enabled capabilities, Atraium can help a room:
- research current information and work from sources;
- inspect, create, and edit local files;
- build software and work through diagnostics;
- create Word documents, PDFs, Excel workbooks, and PowerPoint presentations;
- generate and edit images;
- run local command-line and process workflows;
- work with Visual Studio solutions on supported Windows environments;
- extend its capabilities through plugins, MCP servers, and computer-use handoff.
- use a familiar coding-agent tool surface for file work, search, patching, and shell commands when the Room enables it.
Atraium includes a compact Coding tool source for models that are already familiar with coding-agent workflows. When enabled for a Room, it exposes direct tools for reading, writing, exact edits, globbing, grep-style search, shell commands, and structured patches.
The core uses the familiar names Read, Write, Edit, Glob, Grep, Bash, and apply_patch. It also recognizes common indirect aliases such as read_file, execute_command, exec, search_files, and apply_diff through the local tool gateway.
For routine project commands, Bash accepts a normal command and infers a safe accessible working directory from referenced files or projects. Use cwd only when an explicit override is needed. Specialist Atraium tools remain available for structured text edits, C#, JSON, YAML, XML, Visual Studio, and document work. See the 0.2.3 Preview 1 release notes for details.
A room is more than a conversation thread. Each room can have its own:
- name and image;
- assistant name;
- personality or complete system prompt;
- workflow focus;
- tools and capability mix;
- conversation and memory context.
One room might be a careful research analyst. Another might be a direct developer partner, document studio, presentation creator, image lab, or local operations assistant.
Atraium 0.2.3 Preview 1 uses reusable Model Provider Profiles, so one configured provider target can be assigned to each compatible capability.
- OpenAI-compatible cloud and self-hosted endpoints are supported for their compatible capabilities.
- Anthropic Claude is available for Chat and Utility workflows.
- Google Gemini is available across its supported Chat, Utility, memory, image, realtime, and voice capability set.
- Open-weight models can be used through a compatible OpenAI-style endpoint. Examples include Qwen 3.5 and Qwen 3.6 variants, plus KAT-Coder V2.5 Dev and compatible GGUF builds.
Atraium connects to the API exposed by your serving stack. Serve the model through an OpenAI-compatible endpoint, then create a Model Provider Profile with OpenAI-compatible as the provider and Chat Completions as the API mode.
For a self-hosted model, use the endpoint URL, model identifier, and authentication behavior reported by the server. Verify Chat Completions and tool-calling support with a low-risk Room before enabling broad local access.
Some open-weight models, quantized builds, server implementations, or chat templates accept a system-role message only in the first position. For those profiles, open Chat Completions context caching and enable Update system message with current-turn context. Atraium then keeps all dynamic turn context in the single root system message instead of adding a later system message. The setting is off by default because the stable-root mode is usually better for provider prompt-cache reuse.
KAT-Coder example: enable this switch for the relevant KAT-Coder V2.5 Chat Completions profile when using a build that requires the system message to stay first, including the Abiray KAT-Coder-V2.5-Dev-Imatrix-GGUF release when served through a compatible endpoint. GGUF is a model format and quantization; actual compatibility still depends on the selected server and chat template.
Provider availability, model behavior, regions, quotas, API support, and tool reliability can differ by account and serving stack. See the 0.2.3 Preview 1 release notes for the current capability overview.
The minimum configuration is Chat.
For the best first experience, configure:
- Chat - powers room conversations.
- Utility - supports general assistant workflows.
- Search / Grounding - enables current research and source-backed answers.
Add embeddings, images, realtime, plugins, and MCP servers only when they become useful.
Atraium does not include:
- a model-provider account;
- an API key;
- model credits or token allowance;
- protection from provider charges.
You configure your own supported provider accounts and endpoints. You remain responsible for their usage and costs.
Tool-rich workflows can consume substantially more tokens than a short chat. A single request may involve multiple model calls, tool-selection rounds, search or grounding charges, embeddings, images, retries, and continued work. The visible length of your prompt is not a reliable estimate of final cost.
Before using a paid endpoint:
- configure provider-side budgets, quotas, hard limits, and alerts where available;
- consider a separate provider project or API key for Atraium;
- monitor provider dashboards during long workflows;
- stop workflows that loop or repeatedly fail;
- do not leave expensive workflows unattended.
Provider-side records are authoritative. Atraium and its maintainer cannot reverse, refund, or credit charges made by a provider.
Read the complete Token usage and provider billing policy.
Atraium can work with real files, processes, applications, and desktop workflows. Start with low-risk tasks and give each room only the access it needs.
Review important operations involving:
- sensitive files or source code;
- destructive file changes;
- local processes and applications;
- external services and plugins;
- long-running or high-cost model workflows.
Read the privacy notes and security policy before using the preview with sensitive material.
Preview feedback is welcome, especially about:
- installation and first-run setup;
- provider configuration;
- room creation and instructions;
- workflows that succeeded or failed;
- unclear error messages;
- local-tool behaviour;
- features you would genuinely use again.
Use GitHub Issues for reproducible bugs and structured feedback. Use GitHub Discussions for broader questions and workflow ideas once Discussions is enabled.
Never post API keys, tokens, private endpoint URLs, customer information, confidential prompts, source code, or sensitive local files in a public issue.
Atraium is a passion project and this is an early preview, not a finished commercial product.
Expect:
- rough edges;
- incomplete diagnostics;
- breaking changes between previews;
- best-effort support;
- features that depend heavily on the selected model and provider;
- platform-specific limitations.
Read Preview scope and expectations and Support before reporting an issue.
The central idea remains simple:
You choose the model. You decide how it should behave. You give it the tools it needs. Then you put it to work.