MiraIA

MiraIA is a personal Chrome extension prototype that gives GPT eyes and durable browser memory. It observes supported webpages while active, records meaningful browser events, captures settled viewport screenshots, and keeps that evidence in one OpenAI Conversation.

The goal is simple: a user should be able to ask “What happened, and what should I do now?” without manually attaching screenshots or explaining the browser journey again.

The problem

AI guidance often loses contact with the real browser workflow. The user leaves the chat, opens a form or portal, navigates through several pages and tabs, encounters a redirect or validation message, and then has to reconstruct everything manually before the assistant can help again.

This creates context switching, repeated explanations, and advice based on incomplete information.

The solution

MiraIA combines three surfaces:

• a normal GPT-powered chat in the Chrome side panel;
• one editable conversation name representing the user’s longer-running goal;
• a floating webpage suggestion bubble for timely next actions, warnings, and clarifying questions.

While the side panel is open, MiraIA records neutral browser facts such as navigation, tab changes, meaningful control actions, redirects, visible errors, and settled page states. After related activity becomes quiet, it captures the final viewport and appends the ordered evidence and screenshot to the same durable OpenAI Conversation.

GPT interprets the accumulated evidence. The extension records facts but does not invent workflow conclusions.

How it works

1. The user installs MiraIA and connects a personal OpenAI API key.
2. MiraIA creates one OpenAI Conversation and stores its conversation_id locally.
3. Supported browser activity is collected as ordered raw events.
4. A short debounce groups related events into one settled observation.
5. The final page URL, title, safe event metadata, and viewport screenshot are appended to the Conversation.
6. GPT-5.6 analyses the new evidence and returns a complete list of up to three active suggestions.
7. The user can ask a normal question in the side-panel chat using the same accumulated Conversation.
8. Suggestions may point to and highlight a matching visible control, but MiraIA never clicks it.

Key Mini-MVP features

• Chrome Manifest V3 extension
• Chrome side-panel chat
• One durable OpenAI Conversation
• Automatic meaningful browser observations
• Settled viewport screenshots
• Recent and accumulated browser journey context
• Floating suggestion bubble on webpages
• Optional local element highlighting
• Explicit goal and suggestion decisions
• Persistent local UI and retry state across Chrome restarts
• One complete Reset conversation action

Safety and user control

MiraIA may point, explain, scroll, and highlight. It does not autonomously click, upload, sign, pay, confirm, or submit.

The extension deliberately avoids extracting passwords, payment data, authentication tokens, cookies, hidden fields, or full form payloads as structured event metadata. Viewport screenshots are not masked in this personal prototype, so setup must clearly disclose that anything visible on screen may be sent to OpenAI.

The direct browser API-key design is limited to this personal Mini-MVP and is not suitable for a distributable production extension.

What makes MiraIA different

MiraIA does not rebuild a short prompt from a selected window of local history for every request. Browser observations, screenshots, user decisions, visible chat turns, and hidden suggestion analysis all accumulate in one OpenAI Conversation.

This lets GPT reason from the actual sequence of browser evidence while MiraIA remains a lightweight observation, memory, and guidance layer.

Demo goal

The defining demo question is:

What happened, and what should I do now?

A successful demo shows MiraIA answering from the current page and accumulated browser journey without a manual screenshot upload or repeated explanation.

Build Week status

This Devpost project is a preparation draft. The final submission must be updated to describe only the functionality that is actually implemented and demonstrated. The repository URL, testing instructions, narrated demo video, and primary Codex Session ID will be added before submission.

Built With

• chrome
• chrome-side-panel
• codex
• gpt-5.6
• manifest-v3
• openai-conversations-api
• openai-responses-api
• structured-outputs