diff --git a/Dockerfile b/Dockerfile
index 414260fb681..99c23432b21 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -206,6 +206,10 @@ ARG NEMOCLAW_INFERENCE_API=openai-completions
 ARG NEMOCLAW_CONTEXT_WINDOW=131072
 ARG NEMOCLAW_MAX_TOKENS=4096
 ARG NEMOCLAW_REASONING=false
+# Comma-separated list of input modalities accepted by the primary model
+# (e.g. "text" or "text,image" for vision-capable models). OpenClaw's
+# model schema currently accepts "text" and "image". See #2421.
+ARG NEMOCLAW_INFERENCE_INPUTS=text
 # Per-request inference timeout (seconds) baked into agents.defaults.timeoutSeconds.
 # Increase for slow local inference (e.g., CPU Ollama). openclaw.json is
 # immutable at runtime (Landlock read-only), so this can only be changed by
@@ -255,6 +259,7 @@ ENV NEMOCLAW_MODEL=${NEMOCLAW_MODEL} \
     NEMOCLAW_CONTEXT_WINDOW=${NEMOCLAW_CONTEXT_WINDOW} \
     NEMOCLAW_MAX_TOKENS=${NEMOCLAW_MAX_TOKENS} \
     NEMOCLAW_REASONING=${NEMOCLAW_REASONING} \
+    NEMOCLAW_INFERENCE_INPUTS=${NEMOCLAW_INFERENCE_INPUTS} \
     NEMOCLAW_AGENT_TIMEOUT=${NEMOCLAW_AGENT_TIMEOUT} \
     NEMOCLAW_INFERENCE_COMPAT_B64=${NEMOCLAW_INFERENCE_COMPAT_B64} \
     NEMOCLAW_MESSAGING_CHANNELS_B64=${NEMOCLAW_MESSAGING_CHANNELS_B64} \
@@ -304,6 +309,7 @@ inference_api = os.environ['NEMOCLAW_INFERENCE_API']; \
 context_window = int(os.environ.get('NEMOCLAW_CONTEXT_WINDOW', '131072')); \
 max_tokens = int(os.environ.get('NEMOCLAW_MAX_TOKENS', '4096')); \
 reasoning = os.environ.get('NEMOCLAW_REASONING', 'false') == 'true'; \
+inference_inputs = [v.strip() for v in os.environ.get('NEMOCLAW_INFERENCE_INPUTS', 'text').split(',') if v.strip()] or ['text']; \
 _raw_agent_timeout = os.environ.get('NEMOCLAW_AGENT_TIMEOUT', '600'); \
 agent_timeout = int(_raw_agent_timeout) if _raw_agent_timeout.isdigit() and int(_raw_agent_timeout) > 0 else (_ for _ in ()).throw(ValueError('NEMOCLAW_AGENT_TIMEOUT must be a positive integer')); \
 inference_compat = json.loads(base64.b64decode(os.environ['NEMOCLAW_INFERENCE_COMPAT_B64']).decode('utf-8')); \
@@ -325,7 +331,7 @@ providers = { \
         'baseUrl': inference_base_url, \
         'apiKey': 'unused', \
         'api': inference_api, \
-        'models': [{**({'compat': inference_compat} if inference_compat else {}), 'id': model, 'name': primary_model_ref, 'reasoning': reasoning, 'input': ['text'], 'cost': {'input': 0, 'output': 0, 'cacheRead': 0, 'cacheWrite': 0}, 'contextWindow': context_window, 'maxTokens': max_tokens}] \
+        'models': [{**({'compat': inference_compat} if inference_compat else {}), 'id': model, 'name': primary_model_ref, 'reasoning': reasoning, 'input': inference_inputs, 'cost': {'input': 0, 'output': 0, 'cacheRead': 0, 'cacheWrite': 0}, 'contextWindow': context_window, 'maxTokens': max_tokens}] \
     } \
 }; \
 config = { \
diff --git a/src/lib/onboard.ts b/src/lib/onboard.ts
index 8567fe81446..fd93bf95c9c 100644
--- a/src/lib/onboard.ts
+++ b/src/lib/onboard.ts
@@ -1551,6 +1551,17 @@ function patchStagedDockerfile(
       `ARG NEMOCLAW_REASONING=${reasoning}`,
     );
   }
+  // Honor NEMOCLAW_INFERENCE_INPUTS for vision-capable models. OpenClaw's
+  // model schema currently accepts "text" and "image" only, so validate
+  // strictly against that vocabulary. Adding modalities to OpenClaw later
+  // only requires widening this regex. See #2421.
+  const inferenceInputs = process.env.NEMOCLAW_INFERENCE_INPUTS;
+  if (inferenceInputs && /^(text|image)(,(text|image))*$/.test(inferenceInputs)) {
+    dockerfile = dockerfile.replace(
+      /^ARG NEMOCLAW_INFERENCE_INPUTS=.*$/m,
+      `ARG NEMOCLAW_INFERENCE_INPUTS=${inferenceInputs}`,
+    );
+  }
   // NEMOCLAW_AGENT_TIMEOUT — override agents.defaults.timeoutSeconds at build
   // time. Lets users increase the per-request inference timeout without
   // editing the Dockerfile. Ref: issue #2281
diff --git a/test/onboard.test.ts b/test/onboard.test.ts
index 7c6e95f3cfe..3e29b7dcc69 100644
--- a/test/onboard.test.ts
+++ b/test/onboard.test.ts
@@ -776,6 +776,106 @@ describe("onboard helpers", () => {
     }
   });
 
+  it("regression #2421: bakes NEMOCLAW_INFERENCE_INPUTS into the staged Dockerfile when env is set", () => {
+    const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "nemoclaw-onboard-dockerfile-inputs-"));
+    const dockerfilePath = path.join(tmpDir, "Dockerfile");
+    fs.writeFileSync(
+      dockerfilePath,
+      [
+        "ARG NEMOCLAW_MODEL=nvidia/nemotron-3-super-120b-a12b",
+        "ARG NEMOCLAW_PROVIDER_KEY=nvidia",
+        "ARG NEMOCLAW_PRIMARY_MODEL_REF=nvidia/nemotron-3-super-120b-a12b",
+        "ARG CHAT_UI_URL=http://127.0.0.1:18789",
+        "ARG NEMOCLAW_INFERENCE_BASE_URL=https://inference.local/v1",
+        "ARG NEMOCLAW_INFERENCE_API=openai-completions",
+        "ARG NEMOCLAW_INFERENCE_COMPAT_B64=e30=",
+        "ARG NEMOCLAW_WEB_SEARCH_ENABLED=0",
+        "ARG NEMOCLAW_BUILD_ID=default",
+        "ARG NEMOCLAW_INFERENCE_INPUTS=text",
+      ].join("\n"),
+    );
+
+    const prior = process.env.NEMOCLAW_INFERENCE_INPUTS;
+    process.env.NEMOCLAW_INFERENCE_INPUTS = "text,image";
+    try {
+      patchStagedDockerfile(
+        dockerfilePath,
+        "gpt-5.4",
+        "http://127.0.0.1:18789",
+        "build-inputs",
+        "openai-api",
+      );
+      const patched = fs.readFileSync(dockerfilePath, "utf8");
+      assert.match(patched, /^ARG NEMOCLAW_INFERENCE_INPUTS=text,image$/m);
+    } finally {
+      if (prior === undefined) {
+        delete process.env.NEMOCLAW_INFERENCE_INPUTS;
+      } else {
+        process.env.NEMOCLAW_INFERENCE_INPUTS = prior;
+      }
+      fs.rmSync(tmpDir, { recursive: true, force: true });
+    }
+  });
+
+  it("regression #2421: rejects malformed NEMOCLAW_INFERENCE_INPUTS and keeps default", () => {
+    const tmpDir = fs.mkdtempSync(
+      path.join(os.tmpdir(), "nemoclaw-onboard-dockerfile-inputs-bad-"),
+    );
+    const dockerfilePath = path.join(tmpDir, "Dockerfile");
+    const baseDockerfile = [
+      "ARG NEMOCLAW_MODEL=nvidia/nemotron-3-super-120b-a12b",
+      "ARG NEMOCLAW_PROVIDER_KEY=nvidia",
+      "ARG NEMOCLAW_PRIMARY_MODEL_REF=nvidia/nemotron-3-super-120b-a12b",
+      "ARG CHAT_UI_URL=http://127.0.0.1:18789",
+      "ARG NEMOCLAW_INFERENCE_BASE_URL=https://inference.local/v1",
+      "ARG NEMOCLAW_INFERENCE_API=openai-completions",
+      "ARG NEMOCLAW_INFERENCE_COMPAT_B64=e30=",
+      "ARG NEMOCLAW_WEB_SEARCH_ENABLED=0",
+      "ARG NEMOCLAW_BUILD_ID=default",
+      "ARG NEMOCLAW_INFERENCE_INPUTS=text",
+    ].join("\n");
+
+    const prior = process.env.NEMOCLAW_INFERENCE_INPUTS;
+    try {
+      // Cases that must all leave the default untouched.
+      const rejectCases = [
+        undefined,
+        "audio",
+        "text,",
+        "Text,Image",
+        "text, image",
+        'text"\nRUN rm -rf /',
+      ];
+      for (const [index, value] of rejectCases.entries()) {
+        fs.writeFileSync(dockerfilePath, baseDockerfile);
+        if (value === undefined) {
+          delete process.env.NEMOCLAW_INFERENCE_INPUTS;
+        } else {
+          process.env.NEMOCLAW_INFERENCE_INPUTS = value;
+        }
+        patchStagedDockerfile(
+          dockerfilePath,
+          "gpt-5.4",
+          "http://127.0.0.1:18789",
+          `build-inputs-reject-${index}`,
+          "openai-api",
+        );
+        assert.match(
+          fs.readFileSync(dockerfilePath, "utf8"),
+          /^ARG NEMOCLAW_INFERENCE_INPUTS=text$/m,
+          `value="${String(value)}" should not change the ARG default`,
+        );
+      }
+    } finally {
+      if (prior === undefined) {
+        delete process.env.NEMOCLAW_INFERENCE_INPUTS;
+      } else {
+        process.env.NEMOCLAW_INFERENCE_INPUTS = prior;
+      }
+      fs.rmSync(tmpDir, { recursive: true, force: true });
+    }
+  });
+
   it("regression #1409: rejects malformed NEMOCLAW_PROXY_HOST/PORT and keeps defaults", () => {
     const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "nemoclaw-onboard-dockerfile-proxy-bad-"));
     const dockerfilePath = path.join(tmpDir, "Dockerfile");