LM Studio Models

Offload tasks to local models when quality suffices; avoid web/proprietary/high-stakes.

Key Terms

• model: From GET models key; use in chat and optional load.
• lmstudioapi_url: Default http://127.0.0.1:1234 (paths /api/v1/...).
• responseid / previousresponseid: Chat returns responseid; pass as previousresponseid for stateful.
• instanceid: loadedinstances[].id or modelinstanceid; for unload.

Trigger in frontmatter; below = implementation.

Prerequisites

LM Studio 0.4+, server :1234, models on disk; load/unload via API (JIT optional); Node for script (curl ok).

Complete Workflow

Step 0: Preflight

GET <base>/api/v1/models; non-200 or connection error = server not ready.

exec command:"curl -s -o /dev/null -w '%{http_code}' -H 'Authorization: Bearer lmstudio' http://127.0.0.1:1234/api/v1/models"

Step 1: List Models and Check Loaded

exec command:"curl -s -H 'Authorization: Bearer lmstudio' http://127.0.0.1:1234/api/v1/models"

Parse models[] (key, type, loadedinstances, maxcontextlength, capabilities, paramsstring). If a model has loadedinstances.length > 0 and fits task, skip to Step 5; else pick key for chat (and optional load). Note loadedinstances[].id for optional unload.

Step 2: Model Selection

Pick key from GET response; use as model in chat (optional load). Constraints: vision -> capabilities.vision; embedding -> type=embedding; context -> maxcontextlength. Prefer loaded (loaded_instances non-empty), smaller for speed/larger for reasoning; fallback primary. Optional POST load; else JIT on first chat.

Step 3: Load Model (optional)

Optional: POST /api/v1/models/load { model, context_length?, ... }. JIT: first chat loads; explicit load only for specific options.

Step 4: Verify Loaded (optional)

If explicit load: GET models, confirm loadedinstances. If JIT: no verify; first chat returns modelinstanceid, stats.modelloadtimeseconds.

Step 5: Call API

From the skill folder: node scripts/lmstudio-api.mjs &lt;model&gt; '&lt;task&gt;' [options].

exec command:"node scripts/lmstudio-api.mjs <model> '<task>' --temperature=0.7 --max-output-tokens=2000"

Stateful: add --previous-response-id=<responseid>. Curl: POST <base>/api/v1/chat, body model, input, store, temperature, maxoutputtokens; optional previousresponseid. Parse: output (type message) -> content; responseid, modelinstanceid, stats. Script outputs content, modelinstanceid, response_id, usage.

Step 6: Unload (optional)

Optional: POST /api/v1/models/unload { instanceid }. instanceid from loadedinstances[].id or chat modelinstance_id. JIT+TTL auto-unload; explicit when needed.

exec command:"curl -s -X POST http://127.0.0.1:1234/api/v1/models/unload -H 'Content-Type: application/json' -H 'Authorization: Bearer lmstudio' -d '{\"instance_id\": \"<instance_id>\"}'"

Error Handling

• Model not found -> pick another model from GET response.
• API/server errors -> GET models, check URL.
• Invalid output -> retry.
• Memory -> unload or smaller model.
• Unload fails -> instanceid must match loadedinstances[].id.

Examples

GET /api/v1/models, then script with model key and task. Optional unload per Step 6 (instance_id from response or GET).

exec command:"curl -s -H 'Authorization: Bearer lmstudio' http://127.0.0.1:1234/api/v1/models"
exec command:"node scripts/lmstudio-api.mjs meta-llama-3.1-8b-instruct 'Summarize and extract 5 key points' --temperature=0.7 --max-output-tokens=2000"

LM Studio API Details

Helper/API: see Step 5. Output: content, modelinstanceid, responseid, usage. Auth: Bearer lmstudio. List GET /api/v1/models. Load POST /api/v1/models/load (optional). Unload POST /api/v1/models/unload { instanceid }.

Notes

• LM Studio 0.4+.
• JIT (first chat loads; modelloadtimeseconds in stats); stateful (responseid / previousresponseid).