diff --git a/src/commands/create.ts b/src/commands/create.ts
index 25afcdf..8079385 100644
--- a/src/commands/create.ts
+++ b/src/commands/create.ts
@@ -22,6 +22,8 @@ export interface CreateOptions {
   env?: boolean;
   /** What the agent is building; sent as agent_goal attribution when provided. */
   goal?: string;
+  /** The LLM model driving the session; overrides env detection (which usually finds nothing). */
+  model?: string;
   /** Set false (--no-agent-metadata) to omit attribution fields entirely. */
   agentMetadata?: boolean;
   /** Test seam for the public-IP echo lookup. */
@@ -51,6 +53,7 @@ export async function runCreate(options: CreateOptions): Promise<CreateResult> {
   // Experimental agent attribution — remove with lib/agent-metadata.ts.
   if (options.agentMetadata !== false) {
     request.agentMetadata = detectAgentMetadata(process.env, options.goal);
+    if (options.model) request.agentMetadata.agent_llm_model = options.model;
   }
 
   const envPath = join(process.cwd(), '.env');
diff --git a/src/lib/args.ts b/src/lib/args.ts
index 7e21a47..7c381d9 100644
--- a/src/lib/args.ts
+++ b/src/lib/args.ts
@@ -7,9 +7,9 @@ export type ParsedArgs =
   | { kind: 'error'; message: string };
 
 /**
- * Hand-rolled argv parser — the CLI has one command and seven flags, which is
- * not worth a runtime dependency. Supports `--flag value` and `--flag=value`,
- * and `--ip` may repeat.
+ * Hand-rolled argv parser — the CLI has one command and a handful of flags,
+ * which is not worth a runtime dependency. Supports `--flag value` and
+ * `--flag=value`, and `--ip` may repeat.
  */
 export function parseArgs(argv: string[]): ParsedArgs {
   const options: CreateOptions = {};
@@ -61,6 +61,12 @@ export function parseArgs(argv: string[]): ParsedArgs {
         options.goal = v;
         break;
       }
+      case '--model': {
+        const v = takeValue();
+        if (typeof v !== 'string') return { kind: 'error', message: v.error };
+        options.model = v;
+        break;
+      }
       case '--force':
         options.force = true;
         break;
@@ -106,6 +112,9 @@ Options:
                        at creation)
   --goal <text>        what you are building — sent as attribution alongside the
                        auto-detected agent framework and model
+  --model <name>       the LLM model driving this session — sent as attribution.
+                       Only needed when it cannot be auto-detected from the
+                       environment, which is common
   --force              replace an existing CLOUDINARY_URL in ./.env
   --no-env             do not write ./.env — credentials are only printed
   --no-agent-metadata  do not send agent attribution fields
@@ -113,4 +122,8 @@ Options:
   --api-host <url>     override the Cloudinary API host (or set CLOUDINARY_API_HOST)
   -V, --version        print the version
   -h, --help           show this help
+
+If you are an AI agent, pass --goal and --model so the cloud can be attributed
+to what you are building. For example:
+  npx @cloudinary/cloud --goal "auto-crop and deliver product images" --model <your-model-id>
 `;
diff --git a/test/args.test.mjs b/test/args.test.mjs
index ce85dd4..77b7d7e 100644
--- a/test/args.test.mjs
+++ b/test/args.test.mjs
@@ -11,7 +11,7 @@ test('parses all flags, space and equals forms', () => {
   const parsed = parseArgs([
     'create', '--ip', '203.0.113.7', '--ip=requester_ip',
     '--email=dev@example.com', '--force', '--no-env', '--json',
-    '--goal', 'add image optimization', '--no-agent-metadata',
+    '--goal', 'add image optimization', '--model=claude-fable-5', '--no-agent-metadata',
     '--api-host', 'http://localhost:9999',
   ]);
   assert.equal(parsed.kind, 'create');
@@ -22,6 +22,7 @@ test('parses all flags, space and equals forms', () => {
     env: false,
     json: true,
     goal: 'add image optimization',
+    model: 'claude-fable-5',
     agentMetadata: false,
     apiHost: 'http://localhost:9999',
   });
diff --git a/test/create.test.mjs b/test/create.test.mjs
index 0b6dbb3..df1f85f 100644
--- a/test/create.test.mjs
+++ b/test/create.test.mjs
@@ -309,11 +309,22 @@ test('sends detected agent attribution and --goal by default', () =>
     ),
   ));
 
+test('--model overrides env detection for agent_llm_model', () =>
+  withAgentEnv(() =>
+    inTempCwd(() =>
+      withBodyCapture(async (host, bodies) => {
+        await runCreate({ apiHost: host, ipEchoFetch: noEcho, model: 'gpt-6-codex' });
+        assert.equal(bodies[0].agent_llm_model, 'gpt-6-codex');
+        assert.equal(bodies[0].agent_framework, 'claude-code');
+      }),
+    ),
+  ));
+
 test('--no-agent-metadata omits attribution fields entirely', () =>
   withAgentEnv(() =>
     inTempCwd(() =>
       withBodyCapture(async (host, bodies) => {
-        await runCreate({ apiHost: host, ipEchoFetch: noEcho, agentMetadata: false, goal: 'ignored' });
+        await runCreate({ apiHost: host, ipEchoFetch: noEcho, agentMetadata: false, goal: 'ignored', model: 'ignored' });
         assert.equal('agent_framework' in bodies[0], false);
         assert.equal('agent_llm_model' in bodies[0], false);
         assert.equal('agent_goal' in bodies[0], false);