diff --git a/docs/CODEX.md b/docs/CODEX.md
index 8142883..aeba009 100644
--- a/docs/CODEX.md
+++ b/docs/CODEX.md
@@ -62,6 +62,12 @@ disables discovery. `RELAY_DIR` relocates Switchboard state for an isolated test
   or pick a Codex model in the app's card in the menu-bar panel: Switchboard raises the same connect
   card pre-filled with the current grant plus that model, and one tap widens it (Deny changes nothing).
   New connects pre-select every signed-in provider's default, grouped by provider; untick to refuse.
+- **Class grants (opt-in):** an app may ask for a capability class instead of ids —
+  `claude_connect({ requirements: [{ class: "cloud-coding" }] })`. The user still approves concrete models on
+  the card; the providers behind them become the grant's allowed providers, and a new conversation resolves to
+  any enabled model of that class on those providers. That keeps a grant working when a provider's catalog
+  moves (gpt-5.5 → gpt-6-astra) or one provider is offline. Legacy `scope.models` grants are unchanged and
+  never widened. `claude_capabilities().modelInfo[].classes` lists each model's classes.
 - **Global default vs. an app's grant:** if the global `defaultModel` is not granted to an app, that
   app's effective default is the first enabled model in its own grant — discovery and routing agree,
   and the grant is never widened. An explicit request for an ungranted model is still refused.
diff --git a/packages/protocol/src/permissions.ts b/packages/protocol/src/permissions.ts
index 1e2c20c..c5ed49a 100644
--- a/packages/protocol/src/permissions.ts
+++ b/packages/protocol/src/permissions.ts
@@ -52,6 +52,11 @@ export interface Budgets {
 }
 
 /** The grant object stored (authoritatively) in the daemon and mirrored to the extension UI. */
+/** CAPABILITY CLASS of a model — what an app can ask for instead of a provider's model id, so a grant
+ *  survives a provider's catalog moving under it (2026-09-07: Codex gpt-5.5 → gpt-6-astra) and works on
+ *  whichever granted provider is online. A model may satisfy several classes. */
+export type ModelClass = "cloud-coding" | "cloud-vision" | "local-text";
+
 export interface OriginGrant {
   /** Browser-verified origin, e.g. "https://shop.example". The permission key. */
   origin: string;
@@ -59,6 +64,12 @@ export interface OriginGrant {
   mode: ConsentMode;
   /** Model ids this origin may request. Empty = none. Maps to daemon model backends. */
   models: string[];
+  /** CLASS grant (additive; set only when the app opted in via ScopeRequest.requirements): when present,
+   *  ANY enabled, online model of one of these classes on one of `providers` is also allowed — resolved at
+   *  call time. `models` stays the explicit-id set. Absent on legacy grants ⇒ behaviour unchanged. */
+  classes?: ModelClass[];
+  /** Provider ids the user allowed at consent (derived from the approved models' backends). Never widened. */
+  providers?: string[];
   /** USER-chosen default for new conversations, overriding the app's requested model.
    * Existing conversation IDs keep their starting model. Must be one of `models`;
    * cleared if it falls out of grant. undefined = honor the app's request/global preference. */
@@ -78,6 +89,9 @@ export interface OriginGrant {
  *  in the consent popup — the granted scope may be smaller than requested, never larger. */
 export interface ScopeRequest {
   models?: string[];
+  /** Ask for a CLASS of model instead of ids (e.g. [{ class: "cloud-coding" }]). Opting in makes the grant
+   *  class-based (OriginGrant.classes/providers) so it is not bound to one provider's model ids. */
+  requirements?: Array<{ class: ModelClass }>;
   /** Tool names the site wants; the daemon resolves each to an access class and the user
    *  approves per-tool. Omit to request "read-only tools only". */
   tools?: string[];
diff --git a/packages/protocol/src/rpc.ts b/packages/protocol/src/rpc.ts
index 2cb133f..cc84d7d 100644
--- a/packages/protocol/src/rpc.ts
+++ b/packages/protocol/src/rpc.ts
@@ -46,6 +46,8 @@ export interface ModelInfo {
   capabilities: { vision: boolean; agentic: boolean; warmSessions: boolean };
   /** Codex uses broker MCP tools; Claude Code may also use its runtime's tools. */
   toolSource: "broker-mcp" | "claude-code" | "none";
+  /** Capability classes this model satisfies (see permissions ModelClass). */
+  classes?: Array<"cloud-coding" | "cloud-vision" | "local-text">;
 }
 
 /** Provider capabilities, returned by claude_capabilities for feature detection. */
diff --git a/packages/sidekick/src/backends/registry.ts b/packages/sidekick/src/backends/registry.ts
index 61470eb..119cdec 100644
--- a/packages/sidekick/src/backends/registry.ts
+++ b/packages/sidekick/src/backends/registry.ts
@@ -72,6 +72,18 @@ export class BackendRegistry {
     return [...this.modelToBackend.keys()];
   }
 
+  /** Capability classes a model satisfies (protocol ModelClass). Local runners are text-only; a cloud model
+   *  with an agentic tool loop is "cloud-coding"; a vision-capable cloud model is also "cloud-vision". */
+  classesOf(model: string): Array<"cloud-coding" | "cloud-vision" | "local-text"> {
+    const backend = this.modelToBackend.get(model);
+    if (!backend) return [];
+    if (backend.id === "ollama" || /[:/]/.test(model)) return ["local-text"];
+    const out: Array<"cloud-coding" | "cloud-vision" | "local-text"> = [];
+    if (backend.capabilities?.agentic) out.push("cloud-coding");
+    if (backend.capabilities?.vision) out.push("cloud-vision");
+    return out;
+  }
+
   modelInfo(): ModelInfo[] {
     return this.allowedModels().map((id) => {
       const backend = this.modelToBackend.get(id)!;
@@ -83,6 +95,7 @@ export class BackendRegistry {
           warmSessions: backend.capabilities?.warmSessions === true,
         },
         toolSource: backend.id === "claude-code" ? "claude-code" : backend.capabilities?.agentic ? "broker-mcp" : "none",
+        classes: this.classesOf(id),
       };
     });
   }
diff --git a/packages/sidekick/src/backends/routing.test.ts b/packages/sidekick/src/backends/routing.test.ts
index cf33225..2049cd9 100644
--- a/packages/sidekick/src/backends/routing.test.ts
+++ b/packages/sidekick/src/backends/routing.test.ts
@@ -117,6 +117,7 @@ test("apps discover backend features and their own granted default", async () =>
   assert.deepEqual(a.modelInfo.find((m: any) => m.id === "model-a"), {
     id: "model-a", backend: "codex", hosted: false,
     capabilities: { vision: true, agentic: true, warmSessions: true }, toolSource: "broker-mcp",
+    classes: ["cloud-coding", "cloud-vision"],   // slice 5a: discovery advertises capability classes
   });
   assert.deepEqual(a.modelInfo.find((m: any) => m.id === "local-text").capabilities,
     { vision: false, agentic: false, warmSessions: false });
@@ -438,3 +439,49 @@ test("mixed providers: picking an ungranted provider's model raises a one-tap re
   assert.equal(asked, null); assert.equal(r.ok, true); assert.equal(grants.get(app)?.modelOverride, "claude-a");
 });
 
+test("class grants: resolve per granted provider, never to an excluded one, and survive catalog drift", async () => {
+  const dir = mkdtempSync(join(tmpdir(), "sb-class-"));
+  const backends = new BackendRegistry();
+  const codexModels = ["codex-a"];
+  const mk = (id: string, models: () => string[]): ModelBackend => ({ id, capabilities: { vision: true, agentic: true }, healthy: async () => true, listModels: async () => models(), run: async (params) => ({ text: params.model!, usage: { inputTokens: 1, outputTokens: 1 } }) });
+  backends.register(mk("claude-code", () => ["claude-a"])); backends.register(mk("codex", () => codexModels));
+  await backends.refreshModels();
+  const grants = new GrantStore(dir);
+  const mcp = new McpRegistry();
+  const gate = new Gate(grants, new BudgetLedger(), new AuditLog(dir), { requestWriteConsent: async () => false } as any, mcp);
+  const broker = new Broker({ config: { stateDir: dir }, backends, grants, budgets: new BudgetLedger(), audit: new AuditLog(dir), gate, mcp, storage: new StorageStore(dir), sessions: { end() {} } } as any) as any;
+  const b = { maxCallsPerMin: 100, maxTokensPerDay: 100000 };
+  const classy = "https://classy.test", excl = "https://claude-only-class.test", legacy = "https://legacy-ids.test";
+  grants.upsert(classy, { models: ["claude-a", "codex-a"], tools: [], budgets: b, classes: ["cloud-coding"], providers: ["claude-code", "codex"] });
+  grants.upsert(excl,   { models: ["claude-a"],            tools: [], budgets: b, classes: ["cloud-coding"], providers: ["claude-code"] });
+  grants.upsert(legacy, { models: ["claude-a", "codex-a"], tools: [], budgets: b });
+  try {
+    // default on Codex: the class-granted app honours it; the provider-excluded app never leaves Claude
+    writeFileSync(join(process.env.RELAY_DIR!, "models.json"), JSON.stringify({ disabled: [], defaultModel: "codex-a" }));
+    assert.equal((await broker.complete(classy, { prompt: "x", sessionId: "c1" })).model, "codex-a");
+    assert.equal((await broker.complete(excl, { prompt: "x", sessionId: "e1" })).model, "claude-a");
+    await assert.rejects(broker.complete(excl, { prompt: "x", model: "codex-a", sessionId: "e2" }), /grant|scope|not granted/i);
+    // CATALOG DRIFT: Codex retires codex-a and ships codex-b (2026-09-07: gpt-5.5 → gpt-6-astra)
+    codexModels.splice(0, 1, "codex-b"); await backends.refreshModels();
+    writeFileSync(join(process.env.RELAY_DIR!, "models.json"), JSON.stringify({ disabled: [], defaultModel: "codex-b" }));
+    // the CLASS-granted app follows the provider's new model (class × provider still granted) …
+    assert.equal((await broker.complete(classy, { prompt: "x", sessionId: "c2" })).model, "codex-b");
+    assert.equal(grants.allowsModel(classy, "codex-b"), true);
+    // … the LEGACY id-only app is never widened to it: falls back inside its explicit ids
+    assert.equal((await broker.complete(legacy, { prompt: "x", sessionId: "l1" })).model, "claude-a");
+    assert.equal(grants.allowsModel(legacy, "codex-b"), false);
+    await assert.rejects(broker.complete(legacy, { prompt: "x", model: "codex-b", sessionId: "l2" }), /grant|scope|not granted/i);
+    // a user override may now point at a class-allowed model
+    assert.equal(grants.setModelOverride(classy, "codex-b")?.modelOverride, "codex-b");
+    assert.equal(grants.setModelOverride(legacy, "codex-b"), null);
+    // connect(): an app that asks for a CLASS gets classes + providers derived from what the user approved
+    broker.requestConnectConsent = async (_o: string, body: any) => ({ models: body.models.available, tools: [], budgets: body.budgets });
+    const g = await broker.connect("https://asks-class.test", { reason: "t", tools: [], requirements: [{ class: "cloud-coding" }] });
+    assert.deepEqual(g.classes, ["cloud-coding"]);
+    assert.deepEqual([...g.providers].sort(), ["claude-code", "codex"]);
+    assert.ok((await broker.capabilities("https://asks-class.test")).modelInfo.every((m: any) => Array.isArray(m.classes)));
+  } finally {
+    writeFileSync(join(process.env.RELAY_DIR!, "models.json"), JSON.stringify({ disabled: [] }));
+  }
+});
+
diff --git a/packages/sidekick/src/security/grant-store.ts b/packages/sidekick/src/security/grant-store.ts
index fca521b..20881cc 100644
--- a/packages/sidekick/src/security/grant-store.ts
+++ b/packages/sidekick/src/security/grant-store.ts
@@ -1,6 +1,6 @@
 import { existsSync, readFileSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
-import type { OriginGrant, ScopeRequest, ToolGrant } from "@relay/protocol";
+import type { OriginGrant, ScopeRequest, ToolGrant, ModelClass } from "@relay/protocol";
 import { DEFAULT_BUDGETS } from "@relay/protocol";
 
 /**
@@ -35,6 +35,17 @@ export class GrantStore {
     this.load();
   }
 
+  /** CLASS grants (slice 5a) need to know a model's classes + provider; the registry owns that, so the
+   *  broker injects a resolver rather than this store importing backends. Without one, class grants are inert. */
+  private classResolver: ((model: string) => { classes: string[]; provider: string } | null) | null = null;
+  setClassResolver(fn: (model: string) => { classes: string[]; provider: string } | null) { this.classResolver = fn; }
+  /** Is this model allowed by the grant's CLASS × PROVIDER set (ignoring the explicit id list)? */
+  classAllows(g: OriginGrant, model: string): boolean {
+    if (!g.classes?.length || !g.providers?.length || !this.classResolver) return false;
+    const r = this.classResolver(model);
+    return !!r && g.providers.includes(r.provider) && r.classes.some((c) => (g.classes as string[]).includes(c));
+  }
+
   private load() {
     if (!existsSync(this.file)) return;
     try {
@@ -66,13 +77,15 @@ export class GrantStore {
 
   /** Create/replace a grant from the scope the USER approved (already narrowed by the consent
    *  UI). `approvedTools` carries each tool with its daemon-assigned access class. */
-  upsert(origin: string, approved: { models: string[]; tools: ToolGrant[]; budgets: ScopeRequest["budgets"]; contextKinds?: string[]; expiresAt?: number }): OriginGrant {
+  upsert(origin: string, approved: { models: string[]; tools: ToolGrant[]; budgets: ScopeRequest["budgets"]; contextKinds?: string[]; expiresAt?: number; classes?: ModelClass[]; providers?: string[] }): OriginGrant {
     const now = Date.now();
     const prev = this.grants.get(origin);
     const grant: OriginGrant = {
       origin,
       mode: prev?.mode ?? "ask", // preserve the user's chosen trust mode across re-consents; default ask
       models: approved.models,
+      classes: approved.classes?.length ? approved.classes : undefined,
+      providers: approved.providers?.length ? approved.providers : undefined,
       // Preserve the user's model override across re-consents, but only if it's still granted.
       modelOverride: prev?.modelOverride && approved.models.includes(prev.modelOverride) ? prev.modelOverride : undefined,
       tools: approved.tools,
@@ -92,7 +105,7 @@ export class GrantStore {
   setModelOverride(origin: string, model: string | null): OriginGrant | null {
     const g = this.grants.get(origin);
     if (!g) return null;
-    if (model && !g.models.includes(model)) return null; // can only override to a granted model
+    if (model && !this.allowsModel(origin, model)) return null; // can only override to a granted model (id or class×provider)
     g.modelOverride = model ?? undefined;
     g.updatedAt = Date.now();
     this.persist();
@@ -146,9 +159,9 @@ export class GrantStore {
   allowsModel(origin: string, model: string | undefined): boolean {
     const g = this.get(origin);
     if (!g) return false;
-    if (!model) return g.models.length > 0;
+    if (!model) return g.models.length > 0 || !!g.classes?.length;
     const want = canonicalModel(model);
-    return g.models.some((m) => canonicalModel(m) === want);
+    return g.models.some((m) => canonicalModel(m) === want) || this.classAllows(g, model);
   }
 
   toolGrant(origin: string, name: string): ToolGrant | null {
diff --git a/packages/sidekick/src/server.ts b/packages/sidekick/src/server.ts
index 02603a8..1ef43b4 100644
--- a/packages/sidekick/src/server.ts
+++ b/packages/sidekick/src/server.ts
@@ -27,6 +27,7 @@ import type {
   SbBrandResult,
   GuideRunParams,
   GuideResult,
+  ModelClass,
 } from "@relay/protocol";
 import { BYOP_VERSION, BYOPErrorCode, ProviderError, isTabPrincipal, hostOfTabPrincipal, nativePrincipal } from "@relay/protocol";
 import { CONNECTOR_META, connectorIdOf, connectorsInClass, type ConnectorClass } from "@relay/protocol";
@@ -157,7 +158,9 @@ export class Broker implements ConsentPrompter, NativeHandler {
   private modelStateSignature: string | undefined;
 
   private sessionRoutes: SessionRoutes;
-  constructor(private deps: BrokerDeps) { this.sessionRoutes = new SessionRoutes(deps.config.stateDir); }
+  constructor(private deps: BrokerDeps) {
+    // CLASS grants (slice 5a): the grant store checks class×provider through the registry's view of a model.
+    this.deps.grants?.setClassResolver?.((m) => { const b = this.deps.backends.backendFor(m); return b ? { classes: this.deps.backends.classesOf(m), provider: b.id } : null; }); this.sessionRoutes = new SessionRoutes(deps.config.stateDir); }
 
   start() {
     const { host, port, pairingToken } = this.deps.config;
@@ -1058,7 +1061,13 @@ export class Broker implements ConsentPrompter, NativeHandler {
     // = the UI never asked (a scope-upgrade re-consent may ask later).
     const approvedKinds = (approved as unknown as { contextKinds?: unknown }).contextKinds;
     const contextKinds = Array.isArray(approvedKinds) ? approvedKinds.map((k) => String(k)).filter(Boolean) : undefined;
-    const grant = this.deps.grants.upsert(origin, { models: approved.models, tools, budgets: approved.budgets, contextKinds, expiresAt: approved.expiresAt });
+    // CLASS grant (slice 5a) — only when the app opted in with `requirements`. The user still approved concrete
+    // models on the card; the providers behind those become the allowed provider set, and the requested classes
+    // let a NEW conversation resolve to any enabled model of that class on those providers — so the grant is
+    // not bound to one provider's model ids (catalog drift, provider offline). Legacy apps: unchanged.
+    const classes = (requested.requirements ?? []).map((r) => r.class).filter((c): c is ModelClass => c === "cloud-coding" || c === "cloud-vision" || c === "local-text");
+    const providers = classes.length ? [...new Set(approved.models.map((m) => this.deps.backends.backendFor(m)?.id).filter((id): id is string => !!id))] : [];
+    const grant = this.deps.grants.upsert(origin, { models: approved.models, tools, budgets: approved.budgets, contextKinds, expiresAt: approved.expiresAt, classes, providers });
     const selectedModel = (approved as unknown as { modelOverride?: string }).modelOverride;
     if (selectedModel && approved.models.includes(selectedModel)) this.deps.grants.setModelOverride(origin, selectedModel);
     else if (approved.models.length === 1) this.deps.grants.setModelOverride(origin, approved.models[0]!);
@@ -1431,7 +1440,14 @@ export class Broker implements ConsentPrompter, NativeHandler {
       if (fit.ok) return { ...selected, model: candidate };
       if (explicit) throw this.toolsUnservable(candidate, fit.unservable);
     }
-    const eligible = (this.deps.grants.get(origin)?.models ?? []).filter((m) => this.deps.backends.isAllowed(m) && this.deps.backends.capabilityModels().includes(m));
+    const grantNow = this.deps.grants.get(origin);
+    const idEligible = (grantNow?.models ?? []).filter((m) => this.deps.backends.isAllowed(m) && this.deps.backends.capabilityModels().includes(m));
+    // CLASS grant (slice 5a): any enabled, online model of a granted class on a granted provider is eligible
+    // too — this is what lets a grant survive a provider's catalog moving under it.
+    const classEligible = grantNow?.classes?.length
+      ? this.deps.backends.allowedModels().filter((m) => this.deps.backends.capabilityModels().includes(m) && !idEligible.includes(m) && this.deps.grants.classAllows(grantNow, m))
+      : [];
+    const eligible = [...idEligible, ...classEligible];
     const fitModel = eligible.find((m) => this.toolFit(origin, m, params).ok);
     if (fitModel) return { ...selected, model: fitModel };
     const blamed = candidate ?? eligible[0];