MCP server exposing local LM Studio (Bionic) models on :1234 as typed tools
Two tools with explicit Zod schemas and detailed descriptions. 'respond' has comprehensive parameter documentation (prompt, model, system, max_tokens, temperature, thinking, json_schema, image_path, timeout_s) with type constraints and defaults. 'models' is simpler but functional. Descriptions are substantive (194 - 1024 chars baseline met). However, output schemas are not formally documented, responses are inferred from code comments rather than explicit schema declarations. Error handling is present but recovery guidance is minimal. Tool naming is clear (verb_noun pattern). No security issues detected (no secrets in params). Composition is sound: two independent, single-purpose tools.
List the local LM Studio (Bionic) models: live load state plus the measured capability card (quality /10, tok/s, role, warnings) and the routing rules. Call this when deciding what to delegate locally.
Run a prompt on the local LM Studio (Bionic) models. Free local inference. Model cheat sheet (measured 2026-07-30): qwen/qwen3.5-4b = volume/batch work, 8/10 @ 48 tok/s; gemma-4-e4b-it-mlx = general default, 10/10 @ 28.5, vision; mistralai_ministral-3-14b-instruct-2512-mlx = hardest problems, 10/10 @ 19.3; qwythos-9b-claude-mythos-5-1m = only choice past ~62K tokens of context (~144K); supergemma4-e4b-abliterated-mlx = uncensored variant, invoke by exact id (text-only). model:'auto' (default) uses the already-loaded model to avoid an 8-30s cold load, else the general default. One model resident at a time. DELEGATION CONTRACT: the prompt must be fully self-contained; the local model cannot see this session, so inline every input, constraint, and output-format requirement. Keep on the hosted model instead: tasks needing session context, files, or memory.
Output schemas not formally documented. 'respond' returns text + metadata (served model, elapsed time, usage stats) but no explicit schema declaration. 'models' returns live load state + capability card but structure is inferred from code, not documented for LLM consumption.
Error messages lack recovery guidance. Timeout errors mention 'retry or raise timeout_s' but do not suggest alternative tools or fallback strategies. 'No routable chat model' error does not guide the user to load a model or check LM Studio status.
Model selection logic ('auto' mode) is complex and context-dependent (prefers loaded model, skips auto_route:false). This is documented in the description but not exposed as a discoverable parameter or constraint. LLMs may not understand the cold-load penalty or auto-routing rules without explicit guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Parameter 'json_schema' accepts arbitrary objects (z.record(z.any())) with minimal validation. No constraint on schema complexity, depth, or size. LLMs could pass malformed or oversized schemas, causing silent failures or timeouts.
No pagination or result limits documented for 'models' tool. If LM Studio returns hundreds of models, the response could bloat the context window. No guidance on filtering or limiting results.