lm-studio（LM 工作室）

When to Use

User wants to run local models with LM Studio, connect an app to its local server, or debug weak local inference behavior.

Use this for server readiness, model loading, OpenAI-compatible API integration, embeddings, MCP setup, and local-first operating decisions.

Architecture

Memory lives in ~/lm-studio/. If ~/lm-studio/ does not exist, run setup.md. See memory-template.md for structure.

~/lm-studio/
├── memory.md         # Activation, preferred port, known-good defaults
├── server-notes.md   # Reachability checks and server mode notes
├── model-profiles.md # Verified models by workload
└── incidents.md      # Repeated failures and confirmed fixes

Quick Reference

Requirements

• LM Studio or llmster is already installed on the machine.
• curl and jq are available for smoke tests and response inspection.
• lms is optional but preferred for repeatable server and model operations.
• Keep requests local by default; only add remote MCP servers or network exposure when the user explicitly asks.

Core Rules

1. Prove the server is reachable before changing client code

• Use server-workflows.md to confirm the actual port, endpoint reachability, and model visibility.
• "LM Studio is open" is not enough. Require one real request to succeed before touching integration code.

2. Separate downloaded, listed, loaded, and active models

• Use model-lifecycle.md for discovery, loading, unloading, and verification.
• Never assume a downloaded filename, API model id, CLI identifier, and active runtime instance are the same thing.

3. Prefer OpenAI-compatible endpoints for app integration

• Start from api-recipes.md and change only the base URL and model identifier before rewriting an existing client.
• Verify each workload separately: responses, chat/completions, embeddings, or completions.

4. Match model size and context to machine limits

• Treat slow first token, OOM, and context overflow as runtime-fit problems first, not prompt problems first.
• Reduce model size, quantization burden, or context length before escalating complexity.

5. Validate after every runtime change

• After loading a new model, changing context length, or altering server settings, run one end-to-end smoke test.
• Record the known-good combination in memory so the agent can reuse it instead of rediscovering it.

6. Treat MCP as a separate risk layer

• Use mcp-playbooks.md to connect servers, but debug model serving and MCP behavior independently.
• Never install untrusted MCP servers or silently route local data to remote endpoints.

7. Escalate beyond local when the task exceeds the local setup

• LM Studio is strong for privacy-sensitive work, offline execution, extraction, and controlled agent loops.
• For unsupported capabilities or repeated quality failures, say so explicitly and recommend a stronger remote path.

Common Traps

• Assuming port 1234 without checking reachability -> integrations fail even though the app looks healthy.
• Treating GET /v1/models as proof a model is ready -> Just-In-Time listings can appear before a usable runtime is confirmed.
• Reusing cloud model names in local requests -> the client is fine, but the local model identifier is wrong.
• Forcing JSON, tools, or vision on an unverified local model -> failures get blamed on prompts instead of capability mismatch.
• Leaving large models loaded while debugging another issue -> RAM or VRAM pressure hides the real cause.
• Installing random MCP servers -> privacy and system access boundaries disappear quickly.

Security & Privacy

Data that leaves your machine:

• None by default for local localhost server calls.
• Optional model downloads or MCP servers follow the user's explicit configuration, not this skill's default path.

Data that stays local:

• Prompt content sent to the LM Studio server running on the same machine.
• Notes stored in ~/lm-studio/ if the user wants persistent context.

This skill does NOT:

• Assume remote access is safe by default.
• Store secrets in skill memory files.
• Install MCP servers or open network access without explicit user intent.

Related Skills

Install with clawhub install <slug> if user confirms:

• models — Choose models by workload, context budget, and quality tradeoffs.
• api — Shape request payloads, retries, parsing, and integration debugging.
• self-host — Operate local infrastructure with practical reliability and security habits.
• open-router — Escalate from local-first execution to routed cloud models when capability gaps matter.
• docker — Package helper services or MCP servers consistently on the local machine.

Feedback

• If useful: clawhub star lm-studio
• Stay updated: clawhub sync