diff --git a/README.md b/README.md
index 9dec203..21785d2 100644
--- a/README.md
+++ b/README.md
@@ -18,7 +18,8 @@ At startup, the plugin:
 1. connects to the configured LM Studio server, or the documented default at
    `http://127.0.0.1:1234`;
 2. validates `GET /api/v1/models` against the native LM Studio response shape;
-3. adds `llm` records to OpenCode and excludes embedding records;
+3. adds `llm` records to OpenCode and excludes embedding records, or only the
+   models LM Studio currently holds loaded when `options.onlyLoaded` is set;
 4. maps the model key, display name, vision support, and effective context;
 5. uses the active loaded context when present and the model maximum when the
    model is available for on-demand loading;
@@ -123,6 +124,38 @@ The plugin preserves the LM Studio `key` as the model ID and uses
 `display_name` as the OpenCode display name. There are no model-family or
 model-name heuristics.
 
+### Loaded models only — `options.onlyLoaded`
+
+By default every generative record is offered, loaded or not, so LM Studio can
+load a model on demand when it is selected. Set `onlyLoaded` to restrict the
+provider to models LM Studio currently holds in memory:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "plugin": ["opencode-lmstudio"],
+  "provider": {
+    "lmstudio": {
+      "options": { "onlyLoaded": true }
+    }
+  }
+}
+```
+
+The model list then matches `lms ps`, and the generated whitelist shrinks with
+it so unloaded models leave the picker instead of accumulating (#17). Useful
+with a large library, where most entries are downloaded but idle, and on shared
+or remote servers where on-demand loading is not wanted.
+
+Discovery runs when OpenCode starts, so the list reflects what was loaded at
+that moment; load a different model and restart OpenCode to pick it up. If
+nothing is loaded, the plugin generates an empty model list rather than falling
+back to idle entries. The structured discovery log reports `onlyLoaded` and
+`skippedUnloaded` so the filter is visible.
+
+Only the boolean `true` enables it; any other value keeps the default of
+offering every model.
+
 ### Context limits
 
 For an unloaded model, `max_context_length` becomes OpenCode's context limit so
diff --git a/docs/v1-contract.md b/docs/v1-contract.md
index 6aa9bd9..d08c79c 100644
--- a/docs/v1-contract.md
+++ b/docs/v1-contract.md
@@ -42,6 +42,7 @@ Automatic discovery checks only LM Studio's documented default address,
 | `display_name` | `name` | Use the server's display name |
 | `type: "llm"` | chat model | Include |
 | `type: "embedding"` | none | Exclude from the chat provider |
+| `loaded_instances` | model map membership | Include every record by default; include only non-empty ones under `options.onlyLoaded` |
 | `capabilities.vision` | `attachment`, input modalities | Add image input only when true |
 | `capabilities.trained_for_tool_use` | `tool_call`, structured diagnostics | Keep tools enabled; report native or default handling |
 | `max_context_length` | `limit.context` | Use when no instance is loaded |
diff --git a/src/plugin/enhance-config.ts b/src/plugin/enhance-config.ts
index a6a0fa0..d198105 100644
--- a/src/plugin/enhance-config.ts
+++ b/src/plugin/enhance-config.ts
@@ -12,6 +12,7 @@ import {
   discoverModels,
   getLMStudioApiKey,
   isGenerativeModel,
+  isLoadedModel,
   normalizeLMStudioURL,
   toOpenAICompatibleURL,
 } from "../utils/lmstudio-api.ts"
@@ -19,7 +20,9 @@ import {
 export interface EnhanceConfigResult {
   readonly discovered: number
   readonly discoveryPath: string
+  readonly onlyLoaded: boolean
   readonly skippedEmbeddings: number
+  readonly skippedUnloaded: number
   readonly skippedUnsupported: number
   readonly serverURL: string
   readonly toolUse: {
@@ -41,6 +44,16 @@ function getString(value: unknown): string | undefined {
   return typeof value === "string" && value.length > 0 ? value : undefined
 }
 
+/**
+ * Resolve `options.onlyLoaded`, which restricts the provider to models LM
+ * Studio currently holds in memory. Fails open to `false` — the documented
+ * default is every generative record, so an unusable value never silently
+ * empties the model list.
+ */
+export function resolveOnlyLoaded(value: unknown): boolean {
+  return value === true
+}
+
 export function effectiveContextLength(model: LMStudioModel): number {
   const loaded = model.loaded_instances.map((instance) => instance.config.context_length)
   return loaded.length === 0
@@ -129,8 +142,13 @@ export async function enhanceConfig(config: OpenCodeConfig, log: PluginLogger):
     const apiKey = existing ? getLMStudioApiKey(explicitApiKey, serverURL) : detected?.apiKey
     const response = detected?.response ?? await discoverModels(serverURL, { apiKey })
     const generative = response.models.filter(isGenerativeModel)
+    // `onlyLoaded` mirrors what LM Studio is actually serving right now.
+    // Unloaded models stay discoverable by default so LM Studio can load them
+    // on demand; opting in trades that for a list with no idle entries.
+    const onlyLoaded = resolveOnlyLoaded(existing?.options?.onlyLoaded)
+    const offered = onlyLoaded ? generative.filter(isLoadedModel) : generative
     const discoveredModels = Object.fromEntries(
-      generative.map((model) => [model.key, toModelConfig(model)]),
+      offered.map((model) => [model.key, toModelConfig(model)]),
     )
     const previousGenerated = generatedStates.get(config)
     const generatedWhitelist = previousGenerated?.whitelist !== undefined
@@ -159,15 +177,17 @@ export async function enhanceConfig(config: OpenCodeConfig, log: PluginLogger):
     })
 
     const result = {
-      discovered: generative.length,
+      discovered: offered.length,
       discoveryPath: LM_STUDIO_MODELS_PATH,
+      onlyLoaded,
       skippedEmbeddings: response.models.filter((model) => model.type === "embedding").length,
+      skippedUnloaded: generative.length - offered.length,
       skippedUnsupported: response.models.filter((model) => !isGenerativeModel(model) && model.type !== "embedding").length,
       serverURL,
       toolUse: {
-        default: generative.filter((model) => toolUseMode(model) === "default").map((model) => model.key),
-        native: generative.filter((model) => toolUseMode(model) === "native").map((model) => model.key),
-        unknown: generative.filter((model) => toolUseMode(model) === "unknown").map((model) => model.key),
+        default: offered.filter((model) => toolUseMode(model) === "default").map((model) => model.key),
+        native: offered.filter((model) => toolUseMode(model) === "native").map((model) => model.key),
+        unknown: offered.filter((model) => toolUseMode(model) === "unknown").map((model) => model.key),
       },
     }
     await log("info", "Discovered LM Studio models", result)
diff --git a/src/types/index.ts b/src/types/index.ts
index c41d9bb..d1dfb7d 100644
--- a/src/types/index.ts
+++ b/src/types/index.ts
@@ -9,13 +9,20 @@ export const LMStudioLoadedInstanceSchema = z.looseObject({
   }),
 })
 
+/**
+ * Reasoning levels LM Studio is known to publish. Kept as documentation only:
+ * the schema accepts any string so a level added by a newer LM Studio build
+ * cannot fail validation for the whole response.
+ */
+export const KNOWN_REASONING_LEVELS = ["off", "on", "low", "medium", "high"] as const
+
 /** Capabilities reported for a native v1 LLM record. */
 export const LMStudioCapabilitiesSchema = z.looseObject({
   vision: z.boolean(),
   trained_for_tool_use: z.boolean(),
   reasoning: z.looseObject({
-    allowed_options: z.array(z.enum(["off", "on", "low", "medium", "high"])),
-    default: z.enum(["off", "on", "low", "medium", "high"]),
+    allowed_options: z.array(z.string()),
+    default: z.string(),
   }).optional(),
 })
 
diff --git a/src/utils/lmstudio-api.ts b/src/utils/lmstudio-api.ts
index 8ca8104..18e8799 100644
--- a/src/utils/lmstudio-api.ts
+++ b/src/utils/lmstudio-api.ts
@@ -136,6 +136,11 @@ export function isGenerativeModel(model: LMStudioModel): model is LMStudioModel
   return model.type === "llm"
 }
 
+/** Whether LM Studio currently holds at least one running instance of a model. */
+export function isLoadedModel(model: LMStudioModel): boolean {
+  return model.loaded_instances.length > 0
+}
+
 export interface AutoDetectedLMStudio {
   readonly serverURL: string
   readonly apiKey?: string
diff --git a/test/plugin.test.ts b/test/plugin.test.ts
index 4aa1446..7ea4444 100644
--- a/test/plugin.test.ts
+++ b/test/plugin.test.ts
@@ -4,6 +4,7 @@ import { LMStudioPlugin } from "../src/index.ts"
 import {
   effectiveContextLength,
   enhanceConfig,
+  resolveOnlyLoaded,
   toolUseMode,
   toModelConfig,
 } from "../src/plugin/enhance-config.ts"
@@ -35,6 +36,14 @@ function model(overrides: Record<string, unknown> = {}): LMStudioModel {
   } as LMStudioModel
 }
 
+function loaded(key: string, context = 32_768): LMStudioModel {
+  return model({
+    key,
+    display_name: key,
+    loaded_instances: [{ id: `${key}:0`, config: { context_length: context } }],
+  })
+}
+
 function embedding(key: string, loadedContext?: number): LMStudioModel {
   return model({
     type: "embedding",
@@ -98,6 +107,26 @@ describe("LM Studio native API v1", () => {
     )
   })
 
+  it("keeps discovering when a model reports a reasoning level this release predates", async () => {
+    const fetcher = vi.fn(async () => modelsResponse([
+      model({
+        capabilities: {
+          vision: false,
+          trained_for_tool_use: true,
+          reasoning: { allowed_options: ["off", "low", "medium", "xhigh", "on"], default: "xhigh" },
+        },
+      }),
+      model({ key: "publisher/second", display_name: "Second" }),
+    ]))
+
+    const response = await discoverModels("http://127.0.0.1:1234", {
+      fetch: fetcher as typeof fetch,
+    })
+
+    expect(response.models).toHaveLength(2)
+    expect(response.models[0]?.capabilities?.reasoning?.default).toBe("xhigh")
+  })
+
   it("rejects HTTP-200 error bodies instead of treating status as endpoint support", async () => {
     const fetcher = vi.fn(async () => new Response(JSON.stringify({ error: "Unexpected endpoint" }), { status: 200 }))
 
@@ -251,6 +280,97 @@ describe("config enhancement", () => {
     ])
   })
 
+  it("offers every generative model when onlyLoaded is not configured", async () => {
+    vi.stubGlobal("fetch", vi.fn(async () => modelsResponse([
+      loaded("publisher/running"),
+      model({ key: "publisher/idle", display_name: "Idle" }),
+    ])))
+    const value = config()
+
+    const result = await enhanceConfig(value, logger())
+
+    expect(result).toMatchObject({ discovered: 2, onlyLoaded: false, skippedUnloaded: 0 })
+    expect(Object.keys(value.provider?.lmstudio?.models ?? {})).toEqual([
+      "publisher/running",
+      "publisher/idle",
+    ])
+  })
+
+  it("restricts models and the generated whitelist to loaded instances when onlyLoaded is set", async () => {
+    vi.stubGlobal("fetch", vi.fn(async () => modelsResponse([
+      loaded("publisher/running", 8_192),
+      model({ key: "publisher/idle", display_name: "Idle" }),
+      model({ key: "publisher/also-idle", display_name: "Also Idle" }),
+      embedding("embedding/loaded", 1_024),
+    ])))
+    const value = config({
+      provider: { lmstudio: { options: { baseURL: "http://127.0.0.1:1234/v1", onlyLoaded: true } } },
+    })
+    const log = logger()
+
+    const result = await enhanceConfig(value, log)
+
+    expect(result).toMatchObject({
+      discovered: 1,
+      onlyLoaded: true,
+      skippedUnloaded: 2,
+      skippedEmbeddings: 1,
+    })
+    expect(Object.keys(value.provider?.lmstudio?.models ?? {})).toEqual(["publisher/running"])
+    expect(value.provider?.lmstudio?.whitelist).toEqual(["publisher/running"])
+    // The loaded instance's allocation still drives the context limit.
+    expect(value.provider?.lmstudio?.models?.["publisher/running"]?.limit?.context).toBe(8_192)
+    expect(log).toHaveBeenCalledWith(
+      "info",
+      "Discovered LM Studio models",
+      expect.objectContaining({ onlyLoaded: true, skippedUnloaded: 2 }),
+    )
+  })
+
+  it("generates an empty model list rather than idle entries when onlyLoaded finds nothing loaded", async () => {
+    vi.stubGlobal("fetch", vi.fn(async () => modelsResponse([
+      model({ key: "publisher/idle", display_name: "Idle" }),
+    ])))
+    const value = config({
+      provider: { lmstudio: { options: { baseURL: "http://127.0.0.1:1234/v1", onlyLoaded: true } } },
+    })
+
+    const result = await enhanceConfig(value, logger())
+
+    expect(result).toMatchObject({ discovered: 0, onlyLoaded: true, skippedUnloaded: 1 })
+    expect(value.provider?.lmstudio?.models).toEqual({})
+    expect(value.provider?.lmstudio?.whitelist).toEqual([])
+  })
+
+  it("keeps explicit model overrides while onlyLoaded filters discovered records", async () => {
+    vi.stubGlobal("fetch", vi.fn(async () => modelsResponse([
+      loaded("publisher/running"),
+      model({ key: "publisher/idle", display_name: "Idle" }),
+    ])))
+    const value = config({
+      provider: {
+        lmstudio: {
+          options: { baseURL: "http://127.0.0.1:1234/v1", onlyLoaded: true },
+          models: { "publisher/pinned": { name: "Pinned" } },
+        },
+      },
+    })
+
+    await enhanceConfig(value, logger())
+
+    expect(Object.keys(value.provider?.lmstudio?.models ?? {}).sort()).toEqual([
+      "publisher/pinned",
+      "publisher/running",
+    ])
+  })
+
+  it.each([undefined, false, "true", 1, null])(
+    "fails open to every model when onlyLoaded is %o",
+    (value) => {
+      expect(resolveOnlyLoaded(value)).toBe(false)
+    },
+  )
+
   it("replaces stale generated models and whitelist entries on a later config load", async () => {
     const fetcher = vi.fn()
       .mockResolvedValueOnce(modelsResponse([