Atlas runs against LM Studio by pointing at the OpenAI-compatible API it serves on port 1234. Start the server with lms server start, which listens on http://localhost:1234/v1, then add LM Studio in atlas.json as a provider with that baseURL and any placeholder apiKey, because LM Studio ignores the key entirely. LM Studio is an MCP host, not an MCP server, so its own mcp.json at ~/.lmstudio/mcp.json is entirely separate from Atlas's MCP config. The payoff is air-gapped sessions where no source code leaves the laptop.
LM Studio is an MCP host, not an MCP server
LM Studio is an MCP host, not an MCP server, which is the fact that decides your whole configuration. LM Studio's own mcp.json lives at ~/.lmstudio/mcp.json and is entirely separate from Atlas's MCP config, so adding a server there does nothing for Atlas on port 1234.
A host attaches tools to its own chat interface. A server exposes tools to a client. LM Studio does the former, which means the servers listed in ~/.lmstudio/mcp.json belong to LM Studio's own UI and are invisible to Atlas. Atlas keeps its MCP servers in its own configuration and uses LM Studio purely as an inference backend over the OpenAI-compatible API. Two separate MCP configurations coexist on the same laptop, and confusing them is the most common LM Studio setup mistake.
Starting the LM Studio server with lms server start
Start LM Studio's server with lms server start, which listens on http://localhost:1234/v1 and speaks the OpenAI API. The lms CLI ships with LM Studio, and port 1234 is the default, so a single command brings the endpoint that Atlas will point at into existence.
Confirm the endpoint answers before you touch Atlas, because a provider that cannot reach http://localhost:1234/v1 produces errors that look like model problems and are really process problems. If port 1234 is already taken, change it with lms server start --port 1235 and update the Atlas baseURL to match. To keep the server alive across terminal sessions rather than tied to the shell you launched it from, run lms daemon up.
Configuring the LM Studio provider in atlas.json
Add LM Studio in atlas.json as a provider with baseURL http://localhost:1234/v1 and any placeholder apiKey, because LM Studio ignores the key entirely. The apiKey field must be present to satisfy the OpenAI client contract, but its value is never validated by LM Studio.
Do not wire a secret manager into a local server that performs no authentication. Put a literal placeholder in apiKey and move on. The field that carries meaning is baseURL, and it must match the port lms server start is actually listening on. If you changed the port with lms server start --port 1235, the baseURL becomes http://localhost:1235/v1, and a mismatch here is the second most common cause of a connection failure after the server simply not running.
Getting models and confirming the exact model id
Pull models with lms get, then confirm the exact model id Atlas should request by hitting /v1/models. LM Studio's model identifiers do not always match the name shown in the UI, and Atlas must request the id the API actually returns, not the label a human reads.
The /v1/models endpoint is the authority. Query it after lms get finishes and copy the identifier verbatim into the model field Atlas will send. A request for a model id that LM Studio does not recognize fails in a way that is easy to misread as a broken provider configuration, when in fact the baseURL and the placeholder apiKey were fine all along and only the id was wrong. Check /v1/models first whenever a model request is rejected.
Running Atlas air-gapped on LM Studio
Use LM Studio for air-gapped sessions where no source code may leave the laptop. With lms server start listening on http://localhost:1234/v1, every token Atlas generates is produced on your own hardware, and the codebase Atlas reads never crosses a network boundary at all.
This is the reason most teams reach for LM Studio in the first place. Client work under a strict data agreement, regulated code, or a machine with no outbound connectivity all rule out a hosted provider. LM Studio makes the constraint satisfiable rather than disqualifying: Atlas keeps its full agent loop, its MCP servers, and its terminal workflow, and only the inference moves onto the local GPU. Nothing about the way you work in the terminal changes.
Keeping LM Studio running in the background
Keep LM Studio's server alive in the background with lms daemon up, and change the port with lms server start --port 1235 if 1234 is already in use. Both commands come from the lms CLI, and together they turn a foreground process into infrastructure you can rely on.
A server tied to a terminal window dies with that window, which means Atlas loses its provider the moment you close the tab. lms daemon up detaches it. The port flag matters in the same practical way: if something else on your machine already owns 1234, lms server start --port 1235 moves LM Studio out of the way, and you update the Atlas baseURL to http://localhost:1235/v1 so the two agree.
Setup
- 01Start LM Studio's server with lms server start, which listens on http://localhost:1234/v1 and speaks the OpenAI API.
- 02Add it in atlas.json as a provider with baseURL http://localhost:1234/v1 and any placeholder apiKey, because LM Studio ignores the key entirely.
- 03Change the port with lms server start --port 1235 and keep it alive in the background with lms daemon up
- 04Pull models with lms get, then confirm the exact model id Atlas should request by hitting /v1/models
- 05Remember that LM Studio is an MCP host, not an MCP server: its own mcp.json lives at ~/.lmstudio/mcp.json and is entirely separate from Atlas's MCP config.
- 06Use it for air-gapped sessions where no source code may leave the laptop.
Frequently asked questions
- how do I use Atlas with LM Studio
- Start LM Studio's server with lms server start, which listens on http://localhost:1234/v1, then add it in atlas.json as a provider with that baseURL and any placeholder apiKey.
- is LM Studio an MCP server
- No. LM Studio is an MCP host, not an MCP server. Its own mcp.json lives at ~/.lmstudio/mcp.json and is entirely separate from Atlas's MCP config.
- what port does LM Studio serve on
- Port 1234. LM Studio serves an OpenAI-compatible API at http://localhost:1234/v1 once you run lms server start. Change it with lms server start --port 1235.
- how do I keep the LM Studio server running in the background
- Run lms daemon up. That detaches the server from your terminal session so Atlas does not lose its provider when you close the window.
- how do I find the right LM Studio model id for Atlas
- Pull models with lms get, then confirm the exact model id Atlas should request by hitting /v1/models. The API id does not always match the label shown in the UI.
- what apiKey does LM Studio need in atlas.json
- Any placeholder apiKey, because LM Studio ignores the key entirely. The field that matters is baseURL, set to http://localhost:1234/v1.
- can I run an AI coding agent with no code leaving my laptop
- Yes. Use LM Studio for air-gapped sessions where no source code may leave the laptop. Atlas points at http://localhost:1234/v1 and inference runs entirely on your own hardware.
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.
Write Unit Tests for Untested Code with Atlas in 2026
How to write unit tests for untested code with Atlas in 2026: the lsp tool enumerates exported symbols, grep copies repo conventions, and bash actually runs the suite.
Atlas vs. Goose: Choosing Your AI Coding Agent in 2026
Compare Atlas and Goose for 2026. Atlas offers terminal-native TUI and code-specialized features. Goose provides shareable Recipes and 70+ MCP extensions for general agentic workflows.
Atlas vs Warp: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with Warp, a smart terminal with AI Agent Mode, for developers in 2026. Evaluate features, privacy, and workflow.
Atlas with DeepSeek-R1 32B Distill (Ollama): Local Reasoning on a 24GB Card in 2026
DeepSeek-R1 32B Distill (Ollama) is 20GB of weights with a 128K context, the strongest R1 distill that fits a 24GB consumer GPU. Atlas setup and tradeoffs for 2026.
Atlas vs Magic.dev: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent with permission-gated tools and diff review, against Magic.dev's research claims of 100 million token context models in 2026.
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 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.