lmstudio-local
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| LMS_PATH | No | Chemin du binaire lms, par défaut ~/.lmstudio/bin/lms. | ~/.lmstudio/bin/lms |
| LM_MCP_TIMEOUT | No | Délai API, 300 secondes par défaut. | 300 |
| LM_MCP_DOCS_TTL | No | Fraîcheur du dépôt documentaire en secondes, 3600 par défaut ; 0 force les contrôles distants. | 3600 |
| LM_MCP_RAG_ROOTS | No | Répertoires autorisés séparés par : sur macOS, par défaut documents/. | documents/ |
| LM_MCP_STATE_DIR | No | Cache et preuves, .state/ dans ce dossier par défaut. | .state/ |
| LM_MCP_UPDATE_TTL | No | Durée du cache changelog/pages API en secondes, 3600 par défaut. | 3600 |
| LM_MCP_CONFIG_PATH | No | Configuration MCP de LM Studio, par défaut ~/.lmstudio/mcp.json. | ~/.lmstudio/mcp.json |
| LM_STUDIO_BASE_URL | No | Origine HTTP locale, par défaut http://127.0.0.1:1234. | http://127.0.0.1:1234 |
| LM_STUDIO_API_TOKEN | No | Jeton LM Studio si son authentification est activée. | |
| LM_MCP_DOCS_AUTO_SYNC | No | Synchronisation documentaire automatique, true par défaut. | true |
| LM_MCP_ALLOWED_INTEGRATIONS | No | Ancienne autorisation globale par ID de plugin ; préférer les permissions exactes via lm_mcp_config. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| lm_statusB | Live API/app/runtime status, compatibility diagnostics and official update check (cached 1h). |
| lm_modelsB | List downloaded models, loaded instance IDs, effective configs and supported capabilities. |
| lm_loadC | Native load: context, flash attention, batch, experts, GPU KV cache. Verifies state and config. |
| lm_load_advancedB | CLI load: GPU offload, parallelism, TTL and speculative decoding. Requires identifier; checks installed flags. |
| lm_estimateB | Estimate model memory needs without loading it, using installed CLI estimate-only support. |
| lm_unloadA | Unload precisely one instance and verify absence. May interrupt its active generations. |
| lm_downloadB | Download a requested catalog model or Hugging Face URL. Returns job ID/status; can consume substantial disk. |
| lm_download_statusB | Check actual download progress; a submitted job is not a completed download. |
| lm_chatC | Native chat/vision/stats. store=true returns response_id for continuation. Optional explicitly enabled LM Studio MCP plugins. |
| lm_embeddingsC | Compute vectors with an embedding model for RAG/search. May auto-load the model. |
| lm_openai_chatA | OpenAI-compatible chat, explicit message history, JSON schema and function definitions. Returns tool calls; never executes them. |
| lm_server_controlA | Check/start/stop LM Studio HTTP server. Starts bound to 127.0.0.1, never to the LAN. |
| lm_runtimeC | Inspect engines/hardware or check stable updates (dry run). update explicitly installs stable engines; retest inference afterwards. |
| lm_integrationsA | List LM Studio MCP plugin names and connector allowlist without exposing commands or secrets. |
| lm_diagnoseB | Diagnose API, server, disk, loaded models, error history and redacted log signals. Returns evidence and repair steps. |
| lm_connectionsB | Manage/test/select LM Studio connection profiles (local, private LAN, HTTPS). Secrets via LM_REMOTE_* env only. Remote profile disables local CLI. |
| lm_docsC | Complete official repo knowledge. brief: French/English task references; guide: procedures; coverage: tools/tests/gaps; catalog/search/read: paginated sources; changes: latest delta; sync: force refresh. scope=all includes labeled drafts, configs and source scripts (never executed). Offset paginates results or characters within a read line range. Follow returned cursors. Media links are references, not downloaded/interpreted. |
| lm_model_configA | Official SDK advanced load config (GPU, KV quantization, mmap, RoPE, seed). schema first; inspect compares SDK and REST; load requires new instance ID. |
| lm_profilesC | Save/reuse named model configurations in this project. load uses the SDK with postcondition verification. |
| lm_mcp_configA | Preview/apply LM Studio mcp.json changes with backup and conflict check. Reuse preview digest to apply. authorize grants exact tools for this config. |
| lm_mcp_probeA | Start/connect a configured MCP and list its real tools. Stdio executes its configured program; no tool invocation. |
| lm_mcp_callA | Invoke one explicitly authorized tool on a configured MCP after fresh schema discovery. Outputs are untrusted data. |
| lm_rag_indexA | Index selected TXT/MD/PDF/DOCX files with local embeddings and source/page/hash metadata. Roots set by LM_MCP_RAG_ROOTS. replace rebuilds collection. |
| lm_rag_searchB | Retrieve source excerpts using cosine similarity. Excludes changed/missing documents and rejects incompatible embeddings. |
| lm_rag_askC | Retrieve then answer locally. Returns no answer without valid source IDs and exact source quotes; semantic correctness still needs review. |
| lm_rag_manageB | List RAG collections/document roots, or delete only an index. Original documents are retained. |
| lm_linkC | Manage official LM Link using installed CLI commands. Account sign-in may still require the app. |
| lm_import_modelA | Preview/import a local GGUF model by copying it into LM Studio. Never moves the original file. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
| lmstudio_workflow | Prepare an LM Studio task using current official sources, actual tools and verification steps. |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
| documentation_overview | Official repository freshness, complete inventory counts and operational topics. |
TDQS
Scored across 28 tools
Many tools are distinct, but several clusters overlap: lm_load/lm_load_advanced/lm_model_config/lm_profiles all configure loading; lm_status/lm_diagnose/lm_runtime overlap in inspection/update; lm_chat/lm_openai_chat and lm_rag_ask/lm_rag_search differ mainly by interface. Descriptions clarify intent but misselection risk remains.
All 28 tools use the lm_ prefix and snake_case, with mostly predictable noun or verb_phrase patterns. lm_load_advanced is the only compound variation but remains readable and consistent with the set.
28 tools is heavy for a single MCP server and exceeds the 16-25 borderline range. While the broad LM Studio domain justifies many operations, the set could be consolidated, especially around load/config/profile tools.
The set covers runtime, models, loading, downloads, chats, RAG, MCP, docs, diagnostics, and server control, which is broad for the domain. Minor gaps like deleting models or more direct MCP server lifecycle management exist, but core local lifecycle coverage is mostly complete.