Peji

Product Introduction

1. Definition: Peji is an agentic AI browser application for iPhone (iOS 18.6+). It is a technical implementation of an autonomous web agent that operates directly within the native WebKit session on the user's device.
2. Core Value Proposition: Peji exists to automate complex, multi-step web tasks on an iPhone—such as price comparisons, travel research, and document summarization—by using an AI model to directly control browser tabs. Its primary value is delivering hands-free task completion with user oversight, strong on-device data privacy, and transparent step-by-step execution visible in a chat interface.

Main Features

1. Agentic WebKit Automation: The core feature is an AI agent that autonomously navigates, reads, and interacts with real WebKit browser tabs on the iPhone. It uses a tool-calling language model to execute actions like navigate, read, click, type, scroll, and run a script. Every action is logged as a step in the chat, providing full visibility into the automation process.
2. On-Device Privacy Architecture: Chats, uploaded files, browser session data (cookies, history), and settings are stored exclusively on the iPhone. Dictation uses the on-device Whisper model. iCloud backup is an optional, end-to-end encrypted container. When using the built-in Peji provider, the server acts as a stateless relay, forwarding requests to third-party inference services (like inference.net and Runware) configured for zero data retention, and stores only account metadata and token counters.
3. User Control & Confirmation Workflow: Before executing any high-stakes action (send, buy, pay, book, delete, publish), the agent pauses and requires explicit user confirmation via a tap. Users can also stop a run at any checkpoint and resume later. This creates a "human-in-the-loop" safety mechanism for autonomous browsing.
4. Structured Data Visualization: For outputs like price comparisons, schedules, or numerical summaries, Peji can generate inline visualizations. It renders charts (bar, line) or compact HTML tables directly within the chat reply, using static HTML/CSS with no scripting or network calls for security. These visualizations include source links.
5. Routines System: Users can save and name frequently used prompts (e.g., "Weather in Varese," "Price check") as reusable routines. Tapping a routine pre-fills the chat composer with the saved instruction, streamlining repetitive multi-step research tasks.

Problems Solved

1. Pain Point: Manually performing repetitive, multi-website research and data aggregation on a mobile browser is time-consuming and cumbersome. Switching tabs, copying prices, and summarizing information is inefficient.
2. Target Audience: Busy professionals, researchers, shoppers, and travelers who need quick, compiled information from multiple web sources directly on their iPhone. It also appeals to privacy-conscious users wary of cloud-based AI agents accessing their browsing sessions.
3. Use Cases:
   • Comparative Shopping: Automatically finding and comparing product prices and shipping options across three different e-commerce sites.
   • Travel Planning: Researching flight options, hotel prices, and itineraries from multiple travel sites in one automated task.
   • Personalized News Briefing: Summarizing key developments from the front pages of multiple news outlets.
   • Document Analysis: Uploading a PDF and requesting a summary or specific data extraction.
   • Market Research: Compiling data points, prices, or features from competitor websites into a structured table.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros:
     • True On-Device Agent: Operates within the iPhone's local WebKit session, allowing it to potentially use logged-in website states (e.g., shopping carts) that cloud-based agents cannot access.
     • Transparent Execution: Every browser action (open, read, click) is visible in the chat, providing auditability and trust.
     • Strong Privacy Model: Core user data (chats, files, session) never leaves the device; server architecture is designed for minimal data exposure.
     • Proactive Safety: Mandatory user confirmation for risky actions prevents autonomous purchases or sends.
     • Bring-Your-Own-Model (BYOM): Supports any OpenAI-compatible endpoint, giving advanced users full control over their AI provider and costs.
   • Cons:
     • Platform Lock-in: Exclusively for iPhone (iOS 18.6+). No native Android, Windows, or Mac desktop application.
     • Dependent on Model Capabilities: The agent's success and reliability are tied to the underlying LLM's tool-calling and reasoning accuracy. Errors in page interpretation can lead to failed tasks.
     • Potential for Slow/Methodical Execution: Complex tasks with many steps may take longer than a human skimming pages, as each step is processed sequentially by the AI.
     • Free Tier Limitations: The built-in provider has a monthly usage cap, limiting heavy users on the free plan.

2. Key Alternatives & Differentiation:

   • Perplexity AI / Comet: These are primarily AI-powered search engines with a "Browse" mode. They often fetch and process web pages in the cloud, not within the user's local browser session. Peji differentiates by running the agent locally on the device, maintaining the user's actual browser session and cookies, and offering granular step-by-step control.
   • Arc Search / "Browse for Me": Similar to Perplexity, this feature summarizes web pages from a cloud-fetched context. It lacks the deep, interactive automation (clicking, typing, form filling) and the local session integration that Peji provides. Peji is an agent that performs tasks, not just a summarizer.
   • Traditional Browser + Manual Work: The manual alternative involves opening multiple tabs, reading, copying, and pasting information. Peji automates this entire workflow, saving significant time and effort for research-intensive tasks, albeit with a reliance on AI accuracy.

Frequently Asked Questions (FAQ)

1. Is Peji a secure browser for private tasks? Peji is designed with a strong privacy architecture: your chats, uploaded files, and browsing session data are stored locally on your iPhone. When using its built-in service, it uses third-party AI models with zero data retention policies. For maximum privacy, you can use your own OpenAI-compatible endpoint, in which case no data passes through Peji's servers at all.
2. Can Peji make purchases or log into my accounts automatically? Peji can navigate to sites where you are already logged in (as it uses your local WebKit session), and it can fill forms. However, it is configured to always stop and require explicit user confirmation before finalizing any action that sends data, completes a purchase, or makes a payment. It cannot autonomously spend money.
3. What is the difference between Peji's free plan and Pro subscription? The free plan includes all core features: the agentic browser, routines, custom instructions, iCloud backup, and the ability to use your own AI endpoint. The limitation is a monthly usage allowance for the built-in Peji AI provider. The Pro plan simply provides a larger monthly usage allowance for this built-in provider. All other functionality is identical.
4. How does Peji compare to ChatGPT's web browsing feature? While both can browse the web, ChatGPT typically operates in a isolated, cloud-based browser instance without your personal session data. Peji's agent runs directly on your iPhone within your personal WebKit session, can see sites where you're already logged in, and provides a detailed, step-by-step audit trail of every action it takes in a chat interface.
5. What happens if the AI makes a mistake during a task? Since every action is displayed as a step in the chat, users can monitor progress. If the agent gets stuck or acts incorrectly, the user can stop the run, manually take over in the corresponding browser tab, or adjust the initial instruction and try again. The "human-in-the-loop" design allows for real-time intervention.