lmstudio-cli（lmstudio CLI 搜索）

LM Studio CLI

Use this skill when the real job is operating LM Studio itself: confirming whether lms exists, checking whether a local or remote LM Studio server is actually running, discovering exact model IDs, deciding whether the OpenAI-compatible endpoints are enough, and wiring a downstream tool to the correct base URL and model identifier.

Do not use this as a generic local-LLM comparison skill. Route broad provider comparison or platform selection to research/survey work. Route downstream-tool-specific scanning or appsec operation to that tool's skill (for example strix) once LM Studio itself is verified.

When to use this skill

• A user mentions LM Studio, lms, or an LM Studio server directly
• You need to verify whether LM Studio is running locally or on another authorized host
• You need the exact model IDs returned by /v1/models before wiring another tool
• You need to choose between LM Studio's OpenAI-compatible endpoints and its native REST API
• You need to load, inspect, or confirm models before an agent or CLI can use them
• A downstream tool works with OpenAI-compatible endpoints, but the user needs LM Studio-specific setup help
• A remote/headless LM Studio workflow is failing and you need a deterministic verification path

Instructions

Step 1: Identify the operating mode

Classify the request before touching commands:

1. Local native CLI mode — the machine should have LM Studio installed and lms available
2. Remote HTTP mode — you only need to test or consume an authorized LM Studio endpoint
3. LM Studio-native management mode — the user needs model loading / listing / unload behavior that goes beyond generic OpenAI-compatible calls
4. Downstream wiring mode — the user already has LM Studio running and needs to point another tool at it

Use the smallest mode that answers the request.

Step 2: Check whether local lms exists

If local CLI operation is expected, verify it first:

command -v lms
lms --help

If lms is missing, do not hallucinate local CLI output. Continue with remote HTTP verification only if the host/endpoint is explicitly authorized.

Step 3: Verify server status or endpoint reachability

For local native checks:

lms server status
lms server status --json --quiet

For a remote or client-style smoke test:

curl -fsS http://HOST:PORT/v1/models

Optional minimal response test:

curl -fsS http://HOST:PORT/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "MODEL_ID",
    "messages": [{"role": "user", "content": "reply with exactly OK"}],
    "temperature": 0,
    "max_tokens": 8
  }'

If you want a reusable parser instead of copy-pasting curl, use python3 scripts/check_lmstudio_endpoint.py --base-url http://HOST:PORT/v1.

Step 4: Discover exact model identifiers

Do not guess model names.

Use one of these:

lms ls
lms ls --llm
lms ps --json
curl -fsS http://HOST:PORT/v1/models

If the user needs the exact loaded instance or runtime state, prefer lms ps or the native REST API over a generic downstream client view.

Step 5: Escalate to LM Studio-native management only when needed

Use the native REST API or LM Studio-native commands when OpenAI compatibility is not enough:

curl -fsS http://HOST:PORT/api/v1/models
curl -fsS http://HOST:PORT/api/v1/models/load \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "MODEL_KEY",
    "context_length": 262144,
    "echo_load_config": true
  }'

And locally:

lms load MODEL_KEY --identifier my-model
lms ps --json

Use this path for context-length, load-state, or instance-specific questions. Do not treat every integration problem as a native-management problem if /v1/models and /v1/chat/completions already answer the user's need.

Step 6: Wire downstream tools carefully

For tools that expect an OpenAI-compatible server, pass the exact model ID and base URL:

export LLM_API_BASE="http://HOST:PORT/v1"
export STRIX_LLM="openai/MODEL_ID"

Some OpenAI-compatible clients still insist on an API key field even when LM Studio itself does not need a real provider key. When that happens, set the client-required dummy key only in the downstream tool's config, not as a claim about LM Studio authentication.

Step 7: Report the operating facts, not guesses

A good final report should include:

• whether lms was present locally
• whether the server was verified locally, remotely, or both
• exact base URL tested
• exact model IDs discovered
• whether the user needed OpenAI-compatible calls only or LM Studio-native management
• the final env/config snippet or command needed by the downstream tool

Examples

Example 1: Local machine with LM Studio installed

User asks: "Is lms installed and what model is loaded?"

Recommended flow:

command -v lms
lms server status --json --quiet
lms ps --json

Example 2: Remote LM Studio host for another tool

User asks: "Point Strix at my LM Studio box."

Recommended flow:

curl -fsS http://HOST:PORT/v1/models
python3 scripts/check_lmstudio_endpoint.py --base-url http://HOST:PORT/v1
export STRIX_LLM="openai/MODEL_ID"
export LLM_API_BASE="http://HOST:PORT/v1"

Example 3: Need more than OpenAI-compatible smoke tests

User asks: "Load the model with a bigger context length and tell me the effective settings."

Recommended flow:

curl -fsS http://HOST:PORT/api/v1/models
curl -fsS http://HOST:PORT/api/v1/models/load \
  -H 'Content-Type: application/json' \
  -d '{"model":"MODEL_KEY","context_length":262144,"echo_load_config":true}'

Example 4: Headless failure triage

User asks: "lms server start broke on my Linux VM. What should I check first?"

Recommended flow:

• verify lms --help
• run lms server status --json --quiet
• rerun with verbose logging if needed
• separate local daemon/server failure from downstream OpenAI-client failure before editing configs

Best practices

1. Separate local CLI availability from remote HTTP reachability; they are related but not the same fact.
2. Use the exact model IDs returned by LM Studio instead of shortening names by hand.
3. Prefer the OpenAI-compatible endpoints for downstream-tool wiring, and the native REST/CLI surfaces for model-management questions.
4. When a downstream client insists on an API key field, describe it as a client compatibility quirk rather than an LM Studio requirement.
5. Treat remote private-network hosts as sensitive; only probe endpoints the user has authorized.
6. Escalate to load-state or context-length guidance only when the simpler /v1/models path is not enough.
7. Keep the boundary clear: lmstudio-cli verifies and configures LM Studio itself; downstream-tool skills own what happens after the endpoint is working.

References

• references/cli-and-server-basics.md
• references/endpoints-and-escalation.md
• references/downstream-tool-wiring.md
• scripts/check_lmstudio_endpoint.py
• LM Studio CLI docs
• LM Studio OpenAI compatibility docs
• LM Studio REST API docs