MCP server for hot-swapping llama.cpp models in Claude Code sessions
Server has 4 tools with clear verb-noun naming (list_models, get_current_model, swap_model, create_model_config). Descriptions are present and moderately detailed (avg ~150 chars), explaining what each tool does. Input schemas are visible with type definitions. However, parameter descriptions are sparse or missing entirely, most params lack guidance on format, constraints, or valid ranges. Output is unstructured text rather than typed objects. Error handling exists but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint). Schemas are basic JSON with minimal constraint documentation.
Generate a new service config for a llama-server model. Creates a launchd plist (macOS) or systemd unit (Linux) that can be used with swap_model.
Get the currently loaded llama.cpp model.
List available llama.cpp model configurations and their load status.
Swap to a different llama.cpp model. Unloads any currently loaded model, loads the requested one, and waits for the health endpoint to confirm readiness.
Parameter descriptions missing or minimal. 'model' param in swap_model has description 'Alias of the model to load' but lacks guidance on how to discover valid aliases or what happens if invalid. 'model_path' in create_model_config lacks format constraints (absolute path requirement is stated in code but not in schema description).
Output is unstructured plain text (e.g., 'Model loaded and ready') rather than typed JSON objects. LLMs cannot reliably extract structured data for downstream tool calls. No documented output schema.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). swap_model and create_model_config are destructive/write operations but lack explicit marking. LLMs cannot distinguish safe from risky tools.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Error messages are informative but lack recovery guidance. E.g., 'Model not found' lists available models, but doesn't suggest calling list_models first. 'Health check timed out' doesn't explain retry strategy.
create_model_config has 5 parameters with defaults but no enum constraints or range limits. gpu_layers=-1 is special (all layers) but not documented. port defaults to 8000 with no validation range (1-65535).