Privacy & permissions

Local by default.
Shared by action.

WorkBench keeps reusable workspace material in browser extension storage by default. When you invoke an AI-assisted feature, only the material needed for that action can move into your existing ChatGPT session or a dedicated Help tab.

Local-first workspaceAction-based sharingApproval before install
The privacy boundaryYou control the crossing
↗
Your browserPages, files and the context you are working with.
SOURCE
Selected action
W
WorkBenchPrompts, projects, captures, Skills and Blueprint definitions stay local by default.
LOCAL
Only when needed
AI
ChatGPTSelected material is used when the workflow needs AI-assisted work.
ACTION
Permission ≠ transmission. Browser access enables a capability; the feature or action you choose determines whether context is actually sent.
01 · LOCAL

Your WorkBench workspace stays in the browser.

Prompts, Writing settings, projects, notes, captures, Skills and Blueprint definitions are stored in extension storage by default.

02 · EXPLICIT

AI-assisted work uses the context the job needs.

That can include prompt text, selected webpage content, files you attach, conversation evidence, Adaptive Run tasks or Blueprint evidence.

03 · VISIBLE

Downloads and installs happen when you choose them.

Exports are created on request, and Blueprint Feature Packs remain reviewable previews until you approve installation.

What happens to information

Three contexts.
Explicit boundaries.

Most WorkBench activity stays between the extension and ChatGPT. Browser Control adds a separate, optional relay only after you pair a browser and grant site access.

What stays local in the extension

Reusable workspace material is kept in browser extension storage by default.

  • ✓Prompts, Writing settings, projects and notes.
  • ✓Captures, Skills and Blueprint definitions.
  • ✓Other workspace state needed to keep WorkBench available in the browser.
  • ✓Exports are generated only when you request them.

What can be sent to ChatGPT

AI-assisted features can place the context needed for that job into your existing ChatGPT session or a dedicated Help tab.

  • →Prompt text and files you deliberately attach.
  • →Selected page context, screenshots, PDFs, images, archives or QA context.
  • →Selected conversation evidence or Adaptive Run task context.
  • →Blueprint recording evidence, source guidance and page HTML/CSS when available.

Optional Browser Control service

Browser Control uses account-bound pairing and relay to connect an approved Chrome browser with the WorkBench MCP in ChatGPT. It processes the command and browser result needed for the selected action; this page does not make a permanent-storage claim for that service data.

  • →Published MCP connections use account OAuth and a paired browser.
  • →Pairing does not grant access to any website; per-site permission remains required.
  • →When browser access is denied or a session disconnects, the requested action cannot continue until access or the connection is restored.
Real WorkBench data paths

The feature determines what crosses the boundary.

The browser does not become a blanket data source. Each workflow uses the context that corresponds to the action you selected.

↗

Send to ChatGPT

Use a page link, visible screenshot, full-page PDF, page images, archive or page QA context when you explicitly choose that browser-menu action.

PAGE CONTEXT · ON ACTION
⇄

Adaptive Run

The goal, planned tasks, evidence and selected working context can move into ChatGPT as WorkBench executes the multi-step job.

WORK CONTEXT · AS NEEDED
◇

Skill Forge

Only the conversations you select to teach the Skill are used as evidence for the reusable Skill-building workflow.

SELECTED EVIDENCE
□

Feature Blueprint

Build Preview can send the recording, WorkBench source, builder guidance and page HTML/CSS when available to a ChatGPT Help tab. Nothing is installed until you approve it.

REVIEW BEFORE INSTALL
Browser access

Access should match the feature.

The version of Chrome you install is the authoritative source for the exact permission names. This page explains the product purpose behind that access.

WorkBench · permissions

Why WorkBench needs browser access

Permissions support the browser-native actions shown in the product: current-page context, capture and export, local workspace storage, context-menu actions and feature-specific browser workflows.

Workspace storage
Keeps prompts, settings, projects, notes, captures, Skills, Blueprint definitions and related WorkBench state available in the extension.
LOCAL
Current page context
Lets a user-selected action use the page you are currently working on — for example a page link, screenshot, page QA or Blueprint recording.
ON ACTION
Capture & export
Supports visible screenshots, full-page PDFs, page images, archives and other user-requested outputs from the browser.
ON ACTION
Feature Blueprint
Records selected browser interactions and can use a background ChatGPT Help tab for semantic steps. Build Preview can include the recording, WorkBench source, builder guidance and page HTML/CSS when available.
FEATURE
Context menu
Adds Prompt Workbench to the browser menu so Send, Build, Save and export actions are available where the page context already lives.
BROWSER UI
Important: a permission makes a browser capability possible; it does not mean every capability is used on every page. The action you choose determines the context WorkBench uses.
Permission ≠ transmissionBrowser access enables a capability. Sending data is a separate user- or feature-triggered step.
Local-first ≠ local-onlySome workflows deliberately cross into ChatGPT because that is where the AI-assisted work happens.
Preview ≠ installA Blueprint Feature Pack remains a reviewable preview until you explicitly approve it.
Your controls

Keep, send, export or remove — deliberately.

WorkBench is designed around explicit actions rather than invisible workflow changes.

Controls that matter day to day

The most important privacy control is knowing when a boundary changes. WorkBench surfaces those moments in the UI.

1
Choose what enters ChatGPTUse the Send, Adaptive Run, Skill Forge or Blueprint action that matches the task instead of passing the entire workspace by default.
2
Review before installing a Feature PackBlueprint Build Preview is a review stage. Nothing is installed until you approve it.
3
Create exports only when neededPDFs, screenshots, archives and other downloads are generated when you choose the corresponding action.
4
Remove saved WorkBench materialDelete or replace workspace items through WorkBench controls. Browser extension data can also be cleared through browser controls.

Once content reaches ChatGPT

WorkBench is no longer the only system involved. Content you send into ChatGPT is handled according to the settings, account controls and policies that apply to your ChatGPT use.

How long local material remains

Workspace material stays in browser extension storage until you remove it, the browser clears that extension data, or the extension is uninstalled, subject to browser behaviour.

Local-first is a boundary, not a guarantee against loss.

Browser storage can be cleared or removed. Keep separate copies of anything you need to retain independently, especially exported work or important reusable material.

First use & support

From install to a verified passage.

Browser Control can be revoked from its account portal; a disconnected browser or denied site permission fails closed rather than reading or acting on that site. Do not enable a workflow or provide sensitive material if its data path is unclear.

1. Install WorkBenchAdd the approved extension in Chrome, then open ChatGPT.
2. Connect Browser to ChatGPTPair the browser from the WorkBench connection flow and confirm its visible status.
3. Approve site accessChoose the page WorkBench may use; access is not granted to every website.
4. Ask WorkBench to inspectAsk WorkBench to inspect the selected readable page before taking the next action.
5. Verify exact sourceProvide the exact quotation to check against the chosen source.
6. Inspect the resulting highlightRead the result in ChatGPT and inspect the temporary page highlight.

When the first run does not finish

  • MCP disconnected: reconnect the WorkBench MCP, then retry.
  • Browser not paired: complete the pairing flow and confirm the connected browser.
  • Website permission missing: approve the intended site, then select the page again.
  • Wrong or ambiguous target tab: activate the intended tab and retry with one clear source.
  • The page changed or a stale element reference appeared: re-read the page and choose a current target.
  • Source quotation not found: check the exact wording, source and occurrence.
  • PDF without extractable text: use a text-based version or another readable source.
  • Action attempted but completion unconfirmed: inspect the visible page result before continuing.
  • Unsupported browser operations: use the supported controls described on the Extension page.
Browse FAQsHow the browser extension works
Optional analytics

Useful signals.
No work content.

Anonymous analytics help us understand whether features are reliable and useful. They are optional, and they do not change the local-first boundary.

Anonymous analytics are off by default

If you choose to enable them in WorkBench settings, the extension sends only structural usage and reliability events using a random installation ID and a 30-minute session ID. Turning analytics off clears queued events and those analytics IDs from extension storage.

  • ✓Feature, workflow and outcome signals.
  • ✓Coarse counts and duration ranges that help identify reliability issues.
  • ✓Extension version and the surface that initiated the event.

What analytics never include

Analytics never includes prompt text, responses, full URLs, names or email addresses. The same boundary excludes page content, messages, notes, files, titles and conversation identifiers.

  • →No prompts, ChatGPT responses or selected webpage content.
  • →No account identity, browsing history or full URLs.
  • →No attachments, filenames, notes or saved workspace material.

Browser Control is optional

Browser Control remains disconnected until you pair a browser. Enter a one-time pairing code only in the extension settings; published MCP connections use account OAuth and the paired browser. Website access is a separate permission request, so connecting it does not automatically grant access to all websites. Page reading, screenshots and supported browser actions are limited to permitted sites. Debugger-dependent actions are unavailable in this build.