Atlas can run offline. Turn on offline mode with atlas --offline (or ATLAS_OFFLINE=1) and Atlas reaches only your chosen model endpoint and local services in 2026; telemetry, update checks, and config fetches are all blocked. Point it at a local model with Ollama or LM Studio and your code and its metadata never leave your machine. Offline mode is opt-in, so you decide when to enforce the boundary.
What offline mode does
Turn on offline mode with atlas --offline or ATLAS_OFFLINE=1. Once on, in 2026, Atlas reaches only your configured model endpoint and local services; every other outbound call is blocked before it leaves the machine.
The boundary is enforced in one place rather than left to each feature to respect, so it holds for install telemetry, update and entitlement checks, and config fetches alike. A blocked call is treated like an ordinary offline network failure, so Atlas keeps working rather than erroring: the call is skipped, not sent. Offline mode is opt-in, so the default posture stays convenient and you enforce the boundary when the work calls for it.
Running against a local model
To run fully air-gapped, point Atlas at a model on your own machine. It works with any local OpenAI-compatible server: Ollama on port 11434 or LM Studio on port 1234.
Run atlas models --local and Atlas detects the running local server, lists its chat models, and registers them as a provider you can pick like any other. With a local model selected and offline mode on, the only network Atlas touches is your local model server, so the whole loop stays on hardware you control. You bring your own model; Atlas is the agent around it.
What stays on your machine
When you run a local embedder, Atlas builds its code index on your machine and searches by meaning without sending source to any embedding service. Combined with offline mode, that keeps 3 things local: your code, its index, and its embeddings.
Atlas plans and diffs every change before applying it, and every tool call is checked against allow, ask, and deny rules before it runs, so autonomy stays under your control even while offline. The privacy story is about where the data physically goes: with a local model and a local embedder, there is no third-party copy of your source to worry about, because there is no third party in the path.
Telemetry and updates, honestly
By default Atlas sends 1 anonymous install ping (version, platform, channel, and an anonymous install id), which you can turn off any time with ATLAS_DISABLE_TELEMETRY=1. It never sends your code, prompts, or file paths.
Offline mode goes further: it blocks that telemetry ping along with update checks, entitlement checks, and config fetches, so nothing but your model endpoint and local services is reachable. If you need a hard guarantee that Atlas is not talking to anything except the model you chose, offline mode is that guarantee.
Frequently asked questions
- Can Atlas run offline?
- Yes. Turn on offline mode with atlas --offline (or ATLAS_OFFLINE=1). Atlas then reaches only your model endpoint and local services; everything else is blocked.
- Does my code leave my machine in offline mode?
- No. With a local model and a local embedder, your code, its index, and its embeddings stay on your machine, and offline mode blocks other network calls.
- Does Atlas phone home?
- By default Atlas sends an anonymous install ping you can disable with ATLAS_DISABLE_TELEMETRY=1. In offline mode, telemetry, update checks, and config fetches are all blocked.
- What model can I use offline?
- Any local OpenAI-compatible server: Ollama or LM Studio. Run atlas models --local to detect it and register its models.
- Is offline mode the default?
- No. It is opt-in with --offline or ATLAS_OFFLINE=1, so you choose when to enforce the network boundary rather than having it always on.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas with Phi-4 (local via Ollama): The 16K Context Tradeoff in 2026
Phi-4 (local via Ollama) drives Atlas at Free (self-hosted) pricing with a 16K tokens (16,384) window. Setup, the 14B reasoning case, and where it breaks.
Atlas with IBM Granite 3.3 8B (Ollama): the Free Local small_model for 2026
IBM Granite 3.3 8B (Ollama) is a 4.9GB general model with 128K tokens (131,072) of context, Free (self-hosted). Assign it to small_model in Atlas. Setup and limits.
Atlas with Amazon Nova 2 Lite: Cost, Context, and Setup in 2026
Amazon Nova 2 Lite drives Atlas from inside your AWS account at $0.33 / $2.75 per Mtok with a 128K context. Setup, real tradeoffs, and when to pick another model.
Atlas with Mistral Large 3 (2512): Big Diffs, EU Hosted, 2026
Mistral Large 3 (2512) drives Atlas with a 262,144 token context and a matching 262,144 token output at $0.50 / 1M input tokens and $1.50 / 1M output tokens.
Atlas with Grok Build 0.1: xAI's First Coding-Agent Model (2026)
Grok Build 0.1 runs in Atlas at $1 / $2 per Mtok with a 256K context and 256,000 max output tokens. A coding-specialized model, but it is a 0.1 release.
Atlas with Gemma 3 27B Instruct on Amazon Bedrock: Flat Pricing, 202,752 Tokens (2026)
Gemma 3 27B Instruct in Atlas via Amazon Bedrock: $0.12 per Mtok input, $0.2 per Mtok output, a 202,752 token context, open weights, and an 8,192 token output cap.
Atlas with Command R: The $0.15 per Mtok Background Model for 2026
Command R runs Atlas background work at $0.15 per Mtok input and $0.6 per Mtok output, with a 128,000 token context and a 4,000 token output cap on patches.
Atlas with IBM Granite 4.0 H Micro in 2026
IBM Granite 4.0 H Micro in Atlas, 2026: a hybrid Mamba-Transformer model at $0.017/$0.112 per Mtok on Cloudflare Workers AI, the cheapest input in the registry.