Chatgpt

ChatGPT

Send a prompt to ChatGPT via Chrome browser automation and save the response to a markdown file.

This is a convenience wrapper around the ChatGPT web UI, not a robust automation API. It depends on your logged-in Chrome session, the current DOM structure, clipboard state, and the claude-in-chrome MCP extension being connected.

Prerequisites

• The claude-in-chrome MCP extension must be installed and connected to your Claude Code session.
• ChatGPT must be logged in within Chrome.
• Primary tested environment is macOS. Clipboard extraction uses pbpaste. On Linux, use xclip -selection clipboard -o or wl-paste. On Windows, use powershell -NoProfile -Command Get-Clipboard.

Arguments

chatgpt <model> [effort]

Models: instant, thinking, pro

Effort levels:

• instant takes no effort argument.
• thinking accepts low, medium, high, xhigh.
• pro accepts medium and high.
• If the model is thinking or pro and no effort is supplied, default to medium.
• If the model is instant, do not apply an effort.
• If the model is omitted, ask the user which model they want.

Note: Pro's UI default is Extended (high), so selecting pro without explicit effort requires changing the effort pill from Extended to Standard.

Typical response times (vary widely with prompt complexity, output length, and server load — treat as rough guidance):

Limitations

• Text-only prompts. File attachments and images are not supported. If the user needs to send an image or file, inform them this requires manual interaction.
• Text-only responses. The copy button captures markdown text only. Images, canvases, and other non-text artifacts are not captured. If the response includes non-text content, take a screenshot and tell the user.
• macOS clipboard. The primary extraction method uses pbpaste. See Prerequisites for cross-platform alternatives.

Workflow

All browser interactions use mcp__claude-in-chrome__computer actions (e.g., left_click, type, key, wait, screenshot, scroll_to, zoom). Page inspection uses mcp__claude-in-chrome__read_page. JavaScript execution uses mcp__claude-in-chrome__javascript_tool.

1. Open ChatGPT

Navigate to https://chatgpt.com. After the page loads, call read_page with filter: "interactive" to confirm the textbox and model selector button are present.

If they are missing after two retries, take a screenshot — the page may be showing a login screen, CAPTCHA, or error. Report to the user if unrecoverable.

2. Start a new chat

If the URL contains /c/ (an existing conversation) or any assistant messages are visible, click "New chat" in the sidebar first.

Call read_page again and confirm the main prompt textbox is present.

If the user explicitly wants to continue the current conversation, skip this step and keep the current chat and model.

3. Select the model

1. Click the button with accessibility label "Model selector" (top of page).
2. The dropdown shows menu items that are unlabeled in the accessibility tree. Their order from top to bottom is: Instant (1st), Thinking (2nd), Pro (3rd), Configure (4th). Take a screenshot to verify the order for the current account, then click the item at the correct position.
3. If the requested model is not present in the menu, stop and report which models are available.

Do not assume the options are in a fixed order. Use the current visible text or take a screenshot to verify before clicking by position.

For instant, skip effort selection and continue to step 5.

4. Set the effort level

Only for thinking and pro.

1. Call read_page with filter: "interactive". Near the bottom of the input area, find a button whose text contains the model name — this is the effort pill.
2. Click the effort pill to open the effort dropdown.
3. Click the UI label that matches the requested effort (see the mapping table in Arguments).
4. Verify by calling read_page again and confirming the effort pill text matches the expected label.

Expected pill labels after selection:

• Thinking: "Light", "Thinking" (Standard), "Extended thinking", or "Heavy thinking"
• Pro: "Pro" (Standard) or "Extended Pro"

If the UI does not expose the expected option, stop and report what is visible.

5. Send the prompt

Before sending, record the current assistant message count — this is needed for accurate completion detection:

window.__chatgptAssistantCountBefore = document.querySelectorAll('[data-message-author-role="assistant"]').length;

Then:

1. Click the textbox (placeholder "Ask anything").
2. Type the prompt using the computer tool's type action.
3. Press Enter to submit using the computer tool's key action with text Return.

Multi-line prompts: The textbox sends on Enter. For prompts containing newlines, use form_input to set the full value at once (avoids premature submission), or use clipboard paste (pbcopy + Cmd+V keystroke).

Long prompts (>2000 characters): Break the text into chunks and type each chunk separately. The type action may drop characters on very long strings. After populating the textbox, consider taking a screenshot to spot-check the content before submitting.

Take a screenshot to confirm the prompt appears as a user message in the conversation. If the textbox still contains the prompt (submission failed), press Enter again.

6. Wait for completion

Higher effort levels think deeply and responses take a long time. This is expected, not a failure.

Poll by running this JS via javascript_tool:

const msgs = [...document.querySelectorAll('[data-message-author-role="assistant"]')];
const lastMsg = msgs.at(-1);
const stop = document.querySelector('button[aria-label="Stop streaming"]')
  || document.querySelector('button[data-testid="stop-button"]')
  || document.querySelector('button[aria-label="Stop generating"]');
const actions = lastMsg
  ? lastMsg.parentElement?.querySelector('[role="group"][aria-label="Response actions"]')
  : null;
const done = msgs.length > (window.__chatgptAssistantCountBefore ?? 0) && !stop && !!actions;
JSON.stringify({ done, hasStopButton: !!stop, assistantCount: msgs.length, hasActionsOnLast: !!actions });

If done is false, the model is still generating. Take a screenshot periodically to visually confirm progress.

Polling schedule (max wait is a safety timeout — report to user if exceeded):

Between polls, use the computer tool's wait action. The wait action caps at 30 seconds — for longer intervals, chain multiple waits.

If max wait is exceeded, take a screenshot and inspect the page. Check for an error message, a completed response below the viewport (try scrolling down), or a still-generating indicator. Report to the user and ask how to proceed.

7. Extract the response

Extract the response using the "Copy response" button with clipboard validation:

1. Read the clipboard before copy — run pbpaste and store the result as before.
2. Locate the copy button — use JS to scope to the latest response:

   const msgs = [...document.querySelectorAll('[data-message-author-role="assistant"]')];
   const lastMsg = msgs.at(-1);
   const actionsGroup = lastMsg?.parentElement?.querySelector('[role="group"][aria-label="Response actions"]');
   JSON.stringify({ found: !!actionsGroup });
   

   Then call read_page and click the "Copy response" button within that actions group. If the button is not visible, scroll to the bottom of the page first.
3. Wait 1–2 seconds for the clipboard to populate.
4. Read the clipboard — run pbpaste again as after.
5. Validate — if after is empty or identical to before, the copy failed. Click "Copy response" again, wait 2 seconds, and retry once. If still unchanged, report extraction failure to the user.

Always use the "Copy response" button — it preserves markdown formatting. Do not use .innerText for saved output.

8. Save the response

Write the clipboard content to a markdown file. Use a short, descriptive, kebab-case filename with a timestamp:

Example: chatgpt-20260322-transformer-scaling.md

Report the saved file path, the requested model and effort, and a brief summary of the response to the user.

Multi-turn conversations

When the user wants to send a follow-up in the same ChatGPT conversation:

• Do not start a new chat or change the model.
• Record a fresh assistant count.
• Type the follow-up in the textbox and press Enter.
• Wait for completion and extract the new response using steps 6–8.

Troubleshooting

• Page not loaded: Wait 3 seconds, call read_page again.
• Model selector not found: The page may not have fully loaded. Take a screenshot and retry.
• Menu items unlabeled in accessibility tree: Take a screenshot to verify the visible order before clicking by position. After clicking, verify the selected model by checking the effort pill text.
• Effort pill not visible: Only appears for thinking and pro. For instant, there is no effort selector — this is correct. For thinking/pro, wait 2 seconds after model selection, then read_page again.
• Requested model not available: The pro option is only available on accounts that expose it. Report to user.
• Expected effort not available: Report the visible effort options and stop.
• Response seems stuck: Take a screenshot. If the UI shows an error or rate limit message, report it. Long waits are normal for high-effort models.
• "Something went wrong" during generation: Look for a "Regenerate" button. If found, click it to retry, then restart polling from step 6.
• Rate limit or usage cap: ChatGPT may show a banner about reaching a usage cap. Report the exact message to the user and ask whether to retry after a delay.
• Login screen, CAPTCHA, or interstitial: Stop and tell the user what manual action is required in Chrome. Do not attempt to solve CAPTCHAs.
• Copy button not found: Scroll to the bottom of the page first — the response actions are at the end of the last message.
• Clipboard unchanged after copy: Click "Copy response" again, wait 2 seconds, retry. If still unchanged, report failure.
• Canvas or side panel content: Take a screenshot, inform the user, and suggest they extract canvas content manually.