dsh-claude-code-local

DSH's composer defaults to DeepSeek models, but if you already have a local Claude Code CLI, switching providers means juggling raw CLI invocations and pasting streamed output into a terminal—clunky and fragmented. dsh-claude-code-local closes that gap: pick Claude Code from the model selector inside DSH, and conversations stream back into the composer just like a native DeepSeek session.

Under the hood, every message spawns a local claude child process in stream-json mode. Claude Code runs its own agentic loop—calling its built-in tools, honoring its permission mode and max-turns budget—while each stream_event and result frame is translated into DSH chat text, reasoning, and liveness lines in real time. Credentials stay entirely local (~/.claude, ANTHROPIC_API_KEY, OAuth, Bedrock/Vertex); DSH never stores an API key. Permission modes, tool allow/deny lists, maxTurns, working directory, and the model catalog are all hot-reloadable via settings.yaml—no restart required.

Best for developers who already have a working Claude Code installation and want to drive it through DSH's skills, MCP servers, and composer workflow. Keep in mind each request is a fresh stateless CLI call (no cross-request prompt-cache reuse), and DSH's sandbox does not gate Claude Code's tool execution—the CLI's permissionMode is the security boundary. macOS and Linux only.