- Home
- AI Computer Use Agents
- Scalaris Atlas
Scalaris Atlas
What it is
Scalaris Atlas turns a macOS, Linux, or Windows computer into a local host for AI models and serves them to compatible clients. It accepts OpenAI-compatible clients at localhost and exposes chat completion, embedding, and model-list routes. Atlas includes Pulsar, while its model manager lets users install other models. The API supports a strict subset of the OpenAI HTTP surface, returning a typed 501 response for unimplemented endpoints. The maker says models stay on the machine, with no telemetry or phone-home behavior. Network egress is denied by default, and non-loopback binds are refused unless explicitly enabled. Scalaris Toolbox installs and supervises Atlas. Configuration is kept in one TOML file and reloads when that file changes. If GPU memory is exhausted, CPU fallback works only when `--allow-cpu` is enabled. The Free plan is free, with free-forever access and zero egress by default.
For and against
For
- Runs models locally on three desktop platforms
- Provides OpenAI-compatible API routes
- Includes Pulsar and model installation
- Denies network egress by default
- No telemetry or phone-home behavior
Against
- API supports only a strict subset
- Unimplemented endpoints return typed 501
- CPU fallback requires `--allow-cpu`
- Installed through Scalaris Toolbox
The verdict
Atlas provides a local, API-accessible model host with restrictive network defaults and no telemetry. Users should verify that their client needs fit its limited OpenAI-compatible surface.
Who it suits
Atlas suits people using OpenAI-compatible clients who want to run models locally. It is intended for local deployments on macOS, Linux, or Windows.
Plans and prices
Free
Free forever · zero egress by default
Read 8 Oct 2026 · scalarislab.com
On the record
- API compatibility
- Atlas accepts OpenAI-compatible clients at localhost and exposes chat completions, embeddings, and model-list routes.scalarislab.com · read 8 Oct 2026
- API limits
- Atlas exposes a strict subset of the OpenAI HTTP surface, and unimplemented endpoints return a typed 501 response.scalarislab.com · read 8 Oct 2026
- Configuration
- Atlas configuration is stored in one TOML file, and settings reload when that file changes.scalarislab.com · read 8 Oct 2026
- Installation
- Atlas is installed and supervised through Scalaris Toolbox.scalarislab.com · read 8 Oct 2026
- Integration
- The maker says to point an OpenAI-compatible client at localhost:8787 to integrate Atlas.scalarislab.com · read 8 Oct 2026
- Intended users
- Atlas is for people using OpenAI-compatible clients who want to run AI models locally.scalarislab.com · read 8 Oct 2026
- Local processing
- The maker says models are served locally and nothing leaves the machine.scalarislab.com · read 8 Oct 2026
- Model support
- Atlas includes Pulsar and lets users install their own models through its model manager.scalarislab.com · read 8 Oct 2026
- Network controls
- Egress is denied by default, and Atlas refuses non-loopback binds unless the user explicitly changes the setting.scalarislab.com · read 8 Oct 2026
- Resource handling
- When GPU memory is exhausted, Atlas falls back to CPU only if the user enables --allow-cpu.scalarislab.com · read 8 Oct 2026
- Telemetry
- The maker says Atlas has no telemetry and does not phone home.scalarislab.com · read 8 Oct 2026
- What it does
- Atlas turns a computer into a private AI host that serves local models to compatible clients.scalarislab.com · read 8 Oct 2026
