@codmir/agent-mobile-sdk

@codmir/agent-mobile-sdk

Device-as-a-tool — Let AI agents control mobile apps.

The first SDK that lets an AI agent running on your server reach into a user's phone and perform actions: navigate screens, fill forms, press buttons, read content, trigger calls. The inverse of how mobile apps work today.

Install

npm install @codmir/agent-mobile-sdk

Quick Start

Mobile Client (React Native)

import { MobileAgentBridge, ALL_PRESETS } from "@codmir/agent-mobile-sdk/react-native";
import { io } from "socket.io-client";
import { Alert } from "react-native";
import { router } from "expo-router";

const bridge = new MobileAgentBridge({
  serverUrl: "https://your-server.com",
  authToken: "user-jwt-token",
  projectId: "project-123",

  // User sees a prompt before dangerous actions execute
  onApprovalRequired: async (request) => {
    return new Promise((resolve) => {
      Alert.alert(
        "AI wants to act",
        `${request.description}\n\n${JSON.stringify(request.params)}`,
        [
          { text: "Deny", onPress: () => resolve(false) },
          { text: "Allow", onPress: () => resolve(true) },
        ]
      );
    });
  },
});

// Register action handlers
bridge
  .registerAction(ALL_PRESETS[0], async (params) => {
    // NAVIGATE
    router.push(params.route as string);
    return { navigated: true };
  })
  .registerAction(ALL_PRESETS[1], async (params) => {
    // FILL_FIELD — your app state management fills the field
    return { filled: true };
  })
  .registerAction(ALL_PRESETS[3], async () => {
    // READ_SCREEN — return current screen state
    return {
      route: "/project/123/tasks",
      fields: [],
      buttons: [{ id: "create-task", label: "Create Task", enabled: true }],
    };
  });

// Connect to workroom
await bridge.connect(io);

Server Side (NestJS / Node.js)

import {
  createMobileActionTool,
  createReadScreenTool,
} from "@codmir/agent-mobile-sdk/server";

// Register as tools in your agent's tool registry
const tools = [
  createMobileActionTool(),
  createReadScreenTool(),
  // ... your other tools
];

// When the agent calls mobile_action, forward to the connected device
socket.emit("mobile:command", {
  id: crypto.randomUUID(),
  type: "navigate",
  params: { route: "/project/123/tasks" },
  timestamp: Date.now(),
});

// Listen for results
socket.on("mobile:result", (result) => {
  // Feed back to the agent's tool_use response
  console.log(result.success, result.data);
});

Architecture

┌──────────────────┐     Socket.IO      ┌──────────────────┐
│   AI Workroom    │◄──────────────────►│   Mobile App     │
│                  │                     │                  │
│  Agent calls     │  mobile:command     │  Bridge receives │
│  mobile_action   │────────────────────►│  & executes      │
│                  │                     │                  │
│  Agent reads     │  mobile:result      │  Handler returns │
│  tool result     │◄────────────────────│  result          │
│                  │                     │                  │
│                  │  mobile:read_screen │  Screen reader   │
│                  │────────────────────►│  returns state   │
│                  │  mobile:screen_state│                  │
│                  │◄────────────────────│                  │
└──────────────────┘                     └──────────────────┘

Action Types

Action	Danger	Description
navigate	safe	Navigate to a screen
fill_field	safe	Fill a text input
press_button	confirm	Press a button (prompts user)
read_screen	safe	Read current screen state
scroll	safe	Scroll the view
show_notification	safe	Show a local notification
trigger_call	confirm	Join a voice room
set_status	safe	Update online status
custom	varies	App-specific actions

Danger Levels

• safe — Executes immediately, no user prompt
• confirm — Shows approval dialog on device before executing
• dangerous — Requires explicit user confirmation with action details

Custom Actions

bridge.registerAction(
  {
    type: "custom",
    description: "Create a new task in the current project",
    danger: "confirm",
    params: {
      title: { type: "string", required: true },
      priority: { type: "string", enum: ["low", "medium", "high"] },
    },
  },
  async (params) => {
    const task = await createTask(params.title, params.priority);
    return { taskId: task.id };
  }
);

What Can the AI Do?

The user says to the AI workroom:

"Create a task called 'Fix login bug' with high priority in the mobile app"

The agent:

• Calls mobile_read_screen → sees user is on the project dashboard
• Calls mobile_action({ action: "navigate", params: { route: "/tasks/new" } })
• Calls mobile_action({ action: "fill_field", params: { fieldId: "title", value: "Fix login bug" } })
• Calls mobile_action({ action: "select", params: { fieldId: "priority", value: "high" } })
• Calls mobile_action({ action: "press_button", params: { buttonId: "create" } }) → user sees approval prompt → approves
• Task created.

License

Apache-2.0