Choose an AI integration¶
JSAT can connect to a native coding client, use Ollama directly, or let Ollama launch a coding client with a local or cloud model. These are separate execution paths. Choose the tab that matches the process you intend to run.
Native Codex owns the conversation and model. JSAT supplies MCP tools and the $jsat
dispatcher.
Native Claude Code owns the conversation and model. JSAT supplies MCP tools, /jsat, and
/jsat-help.
A PATH-installed OpenCode owns the model and provider. This route does not invoke Ollama's launcher.
JSAT talks directly to a model served by Ollama. No coding-client harness is involved.
Ollama launches Claude Code and supplies one selected local or cloud model to that process.
Ollama launches OpenCode and supplies one selected local or cloud model. A standalone OpenCode installation is not required.
Ollama launches Codex and supplies one selected local or cloud model.
The distinction that prevents duplicate pulls¶
| Route | Conversation host | Who selects the model? | Needs ollama pull? |
|---|---|---|---|
| Native Claude/Codex/OpenCode | Native client | That client/account | No |
| Direct Ollama | JSAT shell | jsat ai use ollama --model ... |
Local models only |
| Client through Ollama | Claude/Codex/OpenCode | Ollama selector or --model |
Local models only |
| Client through Ollama Cloud | Claude/Codex/OpenCode | Ollama selector or cloud model name | No; use ollama signin |
jsat ai use ollama configures the AI provider used by JSAT's own LLM-backed tools.
jsat ollama --tool TOOL configures only the launched client process. One does not silently
replace the other.
Manage all coding clients¶
With no target, lifecycle commands apply to Claude, Codex, and OpenCode:
Interactive clients open in separate terminals. To keep routes isolated, name one client and
one route, for example jsat start claude --via ollama --model qwen2.5:0.5b.