Fully Local LM Studio Agent (No Cloud)

This playbook builds an agent that runs entirely on a local LM Studio provider through the Pi SDK runtime, with lite-tier (FTS-only) memory and no outbound network. No cloud API keys, no outbound calls, no second local tool required. An optional upgrade to journal-tier semantic recall — using LM Studio’s own embeddings endpoint, still no second tool — is documented at the end.

Who this is for

A privacy-focused user who prefers LM Studio’s GUI-driven local server over Ollama’s CLI daemon and wants the whole stack — chat model and recall — to stay on their own machine.

Goal

An agent that runs entirely on a local LM Studio provider via the Pi SDK runtime, with lite-tier memory and no outbound network.

Features used

• runtime.local-providers — register LM Studio and reference it as lmstudio:<model>
• runtime.multi-backend — the Pi SDK bridge that drives the local provider
• memory.lite — FTS-only recall, no embeddings or chat LLM required

Configuration

Every key below is verified against the config blueprint. The providers map registers LM Studio; runtime.model references it as lmstudio:qwen3.6-32b. type: "lmstudio" defaults to baseUrl: "http://localhost:1234" when omitted — LM Studio’s default local-server port — shown explicitly here for clarity.

{
  "runtime": {
    "model": "lmstudio:qwen3.6-32b"
  },
  "providers": {
    "lmstudio": {
      "type": "lmstudio",
      "baseUrl": "http://localhost:1234",
      "enabled": true
    }
  },
  "memory": {
    "mode": "lite",
    "path": "./.mono-agent/memory",
    "writeMode": "append-host-summary"
  },
  "webhook": {
    "enabled": true
  }
}

Note

lite mode needs no memory.embeddings — recall is FTS-only, so the scaffolded recipe has exactly one local dependency (LM Studio) and no required secret, so it validates immediately with an empty .env. See Upgrade to semantic recall below if you want journal-tier hybrid recall — it’s still LM Studio-only, just not the zero-config default.

Steps

1. Open LM Studio, load a model, and start the local server from the Developer tab (default port 1234).
2. Scaffold the agent: mono-agent init --model lmstudio:<your-model-id> --memory lite — a lmstudio:* model auto-adds the LM Studio provider block. (There is no LM Studio preset; the local-private preset is Ollama-based.)
3. If your loaded model’s id differs from qwen3.6-32b, update runtime.model to match exactly what LM Studio reports for the loaded model.
4. Run mono-agent validate to confirm LM Studio’s local server is reachable.
5. Run mono-agent start — keep the webhook channel as the zero-credential smoke channel.
6. curl the webhook endpoint and confirm a response generated by the local model.

Smoke test

Tip

curl -X POST the webhook invoke path with a prompt; verify a response generated by the local LM Studio model and that the run artifact shows no outbound non-localhost network.

Upgrade to semantic recall (journal tier)

LM Studio is a first-class embeddings provider. Journal-tier hybrid (BM25 + vector) recall uses the same local server’s /v1/embeddings endpoint and never sends an Ollama request or requires an OpenAI/dummy key. Add it deliberately after the base agent is working:

1. In LM Studio, load an embedding model too (e.g. a Nomic Embed Text v1.5 GGUF) alongside your chat model — LM Studio serves both from the same local server concurrently.
2. Rerun guided mono-agent init in a fresh folder and choose Journal → LM Studio, or add the equivalent explicit block and switch mode to journal:

   {
     "memory": {
       "mode": "journal",
       "path": "./.mono-agent/memory",
       "embeddings": {
         "provider": "lmstudio",
         "model": "text-embedding-nomic-embed-text-v1.5",
         "endpoint": "http://localhost:1234",
         "dim": 768
       }
     }
   }

3. Keep apiKeyEnv absent for LM Studio’s default keyless server. If you enabled LM Studio authentication, add "apiKeyEnv": "LM_STUDIO_API_KEY" and populate that variable in the agent’s owner-only .env; a declared but missing value reports waiting and does not retry keyless.
4. Run mono-agent memory rebuild --json for an existing Journal/BuJo root, then mono-agent validate. A provider/model/dimension change is a new managed index identity; never reuse or relabel old vectors.

Caution

LM Studio identifies models by service-native keys. Set memory.embeddings.model to the exact key whose /api/v1/models entry has type: "embedding"; do not infer it from the runtime chat catalog. Guided setup filters that typed list and proves one real finite vector through /v1/embeddings. If discovery is unavailable, manual model/dimension entry remains available, but the real readiness probe must still pass before setup reports ready.

Related

• Local providers
• Runtime backends
• Embeddings
• Capture and recall
• Webhook channel
• Artifacts and traces — confirm no outbound network in the run artifact
• Fully Local Ollama Agent — the Ollama-embeddings variant, and an option if you’d rather not load two models in LM Studio at once
• Skills and the mono-agent-composer — scaffold this config end to end