SeaShell is model-agnostic and runs against a local model as easily as a hosted one. Point it at any local OpenAI-compatible server, Ollama or LM Studio, and run seashell models --local: in 2026 SeaShell detects the running server, lists its chat models, and registers them as a provider you pick like any other. A local model runs on your machine with no per-token API cost, and pairs with offline mode to run fully air-gapped.
Why run a local model
There are 3 reasons developers point an AI coding agent at a local model in 2026: keeping code on their own machine, avoiding per-token API bills, and not depending on a hosted provider's uptime or pricing.
A local model turns the cost of heavy agent use from a metered API bill into the fixed cost of your own hardware, which matters when an agent makes many calls per task. It also keeps the model side of the loop on your machine, so paired with a local code index there is no third-party in the path at all. SeaShell stays model-agnostic either way: you can run local, bring your own hosted keys, or mix them per task.
Ollama and LM Studio
Run seashell models --local. SeaShell probes the 2 common local servers, Ollama on port 11434 and LM Studio on port 1234, finds whichever is running, and lists its installed chat models.
It then registers those models as a provider in your config, so they appear in the model picker and in seashell models exactly like a hosted provider. There is no config file to hand-edit and no schema to reverse-engineer. If your server runs on a non-default port, set SEASHELL_LOCAL_MODEL_URL and SeaShell probes that instead. Embedding-only models are filtered out, so the picker shows only models you can actually chat with.
Model-agnostic by design
SeaShell resolves the model per message, so you are never locked to 1 provider. Switch between a local model and a hosted one from the picker, or use different models for different work.
That means a local model is not an all-or-nothing commitment: run a small local model for routine edits and switch to a hosted frontier model for the hard problems, without changing tools. Bring your own keys for the hosted providers you already pay for, and SeaShell routes to whichever you pick. The agent, the planning, the diffs, and the permission gating stay the same regardless of which model is behind them.
Local model plus offline mode
A local model is what makes air-gapped work possible. With a local model selected and offline mode on (seashell --offline), the only network SeaShell touches is your local model server on your own machine.
This is the combination for developers who cannot send code to a cloud at all: the model is local, the code index can use local embeddings, and offline mode blocks every other outbound call. See the offline mode guide for how the network boundary is enforced. You bring the model; SeaShell keeps the rest of the loop on hardware you control.
Frequently asked questions
- Can SeaShell use a local model?
- Yes. SeaShell works with any local OpenAI-compatible server, including Ollama and LM Studio. Run seashell models --local to detect it and register its models.
- How do I connect Ollama to SeaShell?
- Run seashell models --local. SeaShell detects Ollama on port 11434, lists your installed chat models, and registers them as a provider you can pick.
- Does SeaShell work with LM Studio?
- Yes. LM Studio on port 1234 is auto-detected too. For a non-default port, set SEASHELL_LOCAL_MODEL_URL and SeaShell probes that endpoint.
- Do I need an API key for a local model?
- No. A local model runs on your machine with no per-token cost. You can still bring API keys for hosted models and switch between them.
- Can I run fully offline with a local model?
- Yes. Pair a local model with offline mode (seashell --offline) so the only network SeaShell touches is your local model server.
Try SeaShell in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install SeaShellRelated guides
Atlas with Llama 3.3 8B Instruct (Meta Llama API) in 2026
Explore Atlas with Llama 3.3 8B Instruct (Meta Llama API) for your coding agent in 2026. Leverage its 128,000 token context window for cheap context reading, ideal for the small_model slot.
Atlas for Java in 2026
Adopt Atlas, the terminal-native AI coding agent, for Java development in 2026. Enhance your workflow with intelligent code search, refactoring, and robust safety features for Maven and Gradle projects.
Atlas with Ollama Cloud in 2026: direct Scaling for Your AI Coding Agent
In 2026, drive Atlas with Ollama Cloud for direct scaling of open-weights models. Access massive context windows up to 1,048,576 tokens and powerful models like qwen3-coder:480b, maintaining local Ollama continuity.
Atlas for Polars: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Polars. Build LazyFrame chains, push scan_csv predicates into the reader, and read explain() plans in 2026.
Atlas with GPT-4.1 in 2026
Explore Atlas with GPT-4.1, OpenAI's April 2025 non-reasoning workhorse. Leverage its 1,047,576 token context for bulk code comprehension at $2 input / $8 output per Mtok. Ideal for large-scale code analysis in 2026.
Atlas with NVIDIA NIM in 2026
Explore NVIDIA NIM as a model provider for Atlas in 2026. Access Nemotron 3 Ultra 550B A55B with a 1,000,000 token context and cost-effective open-weights.
Atlas vs CodeGPT: Terminal AI Coding Agents in 2026
Atlas and CodeGPT comparison for 2026. Atlas offers a terminal-native TUI and permission-gated tool calls. CodeGPT provides a Knowledge Graph and individual file approval workflow.
Atlas with GLM-4.5-Flash in 2026
Explore GLM-4.5-Flash for Atlas in 2026: a free, reasoning-enabled model with a 128K token context window, ideal for background tasks despite rate limits.