Page Assist on a 24 GB Mac: Sidebar Found, Page Not Understood

'So interesting idea to basically have a sidebar.' He built a sidebar for the PromptQuorum app and finds it useful for connecting to a local model. He also uses the Claude Chrome extension 'always when I am lost'.

Objective in his words: the sidebar, and what it can do.

The README (pasted by the tester) lists prerequisites: Bun, Ollama, or any OpenAI-compatible endpoint such as LM Studio or llamafile. Steps: git clone, bun install, bun run build ('by default it will build for Chrome, Edge and Firefox'), then load the build directory as an unpacked extension. A note says to use npm if Bun causes problems.

'Very straightforward, but not for people who don't like terminals.'

git clone completed (14,361 objects, 9.93 MiB). The first 'bun install' failed with 'zsh: command not found: bun'. The tester asked an AI assistant, which said the Bun install had worked and the terminal had not loaded the updated PATH; the fix was 'exec /bin/zsh'.

After that the install 'took a while', about a minute in the terminal, and the extension ran. The 'load unpacked' step is not described in the notes.

The extension ID in the address bar of the first screenshots (ofcf...) differs from the store version's (jfgf...), which is consistent with an unpacked build. Not confirmed.

Clicking the extension opens 'Page Assist - A Web UI for Local AI Models' as a full browser tab. It shows 'Ollama is running', empty model and prompt pickers and a message box.

'It's sold as a sidebar, but it's not a sidebar.'

The settings menu is long: General, Pipeline, Ollama, Chrome AI, OpenAI Compatible API, MCP Servers, Page Action, WebMCP, Memory, Manage Models, Manage Knowledge, Manage Prompts, Manage Share, About. Several carry a Beta badge. 'Chrome AI Settings' offers a Gemini Nano download of about 4 GB with a toggle that is off.

'The system settings look very impressive.'

The README lists the Chrome Web Store link lower on the page. The tester missed it and took the source route. 'It was listed below and I missed it.'

The install dialog says the extension can 'Read and change all your data on all websites' and 'Display notifications'.

The store listing shows 'Featured', 4.8 stars from 245 ratings and 200,000 users, as seen at the test date. Not verified.

Per the tester, the repository's latest update was 4 October 2026, four days before the test. Not captured in a screenshot.

'This Chrome extension download, so easy compared to GitHub.'

In the store version the tester wrote: 'Okay, I'm principle lost. Can you tell me what page assist actually does?' and explained that he expected a sidebar next to a web page. Model: gemma4:latest. The reply is a collapsed 'Thought for 0ms' box with no answer text.

He asked 'Okay, where's my answer?'. The second reply was again an empty 'Thought for 0ms' box.

The cause is not isolated. The model, the thinking toggles or the connection to Ollama could each play a part. Not tested.

A welcome window with a video appeared. Per the tester, it showed that this is not an extension that opens a sidebar taking over the browser, and said a shortcut is needed to open it. The tester tried Ctrl+Shift+Y. It did not work. The video itself is not in the notes.

He found 'sort of traces of the sidebar in the settings' (a Sidebar Position setting is in the pasted settings list). He could not open it.

The README feature list, pasted by the tester, says: Sidebar ('can be opened on any webpage'), Web UI ('like a ChatGPT website') and Chat With Webpage.

For comparison he showed his own PromptQuorum Optimizer extension: a toolbar popup that was offline ('Start the PromptQuorum desktop app to enable optimization'). The notes call it his sidebar and say 'it works'; the screenshot shows the offline state.

An article on aihubmix.com, found through his Ahrefs alerts, referred to his Page Assist review and gave him the energy to try again. The tester has not said whether he read the article.

He worked with ChatGPT and found that there is no switch for the sidebar. On his Mac in Chrome the side panel opens from a right-click on the page; clicking the extension icon opens the full-screen UI, 'a totally different function'.

The settings list also has two Beta entries, 'Set Default Action for Extension Icon Clicks' and 'Set Default action for Context Menu'. They may control this behavior. Not tested.

The first question in the side panel was 'what does the text talk about'. Model shown: gemma4:latest. No answer appeared, only a cursor.

The tester compares: his own Chrome app checks every 20 seconds whether the local model is reachable. Here nothing told him whether the model was working. 'Now I need to find out if the local LLM is working or what the issue is.'

After 'some more clicks' it worked. The notes do not say what fixed it. Later screenshots show qwen3-coder:30b selected instead of gemma4:latest.

Switching the model may have been the fix. Not stated in the notes.

The start screen shows 'Chat with current page' ticked, and 'Page Action' and 'Page tools (WebMCP)' unticked.

He asked it to count words on the page. 'It counted well.' Which page and which model: not recorded.

On the Search Console performance page the reply said the provided context did not contain any CTR information and described the content as a graphical representation of time intervals on a timeline (top of this page). The page shows an Average CTR card of 1.5%. The tester summarizes: 'could not read data, did not understand it'.

Dynamic chart and card content may not reach the model as text. A small context window may also play a part. Neither was tested.

He asked it to find instructions on a website. 'And no. Failed.'

'If you load a new page, the system does not understand it and still works on the old page.' He calls it a refresh issue and concludes that jumping from page to page is not possible. Number of tries and pages used: not recorded; not captured in a screenshot.

He found a setting that suggests the tool can click, search and fill in data (Page Action). 'Nothing like this': he could not get it even to scroll the page.

Whether Page Action or WebMCP was switched on during that test is not recorded. In the screenshot both are off. If they were off, this is a setup miss rather than a product limit.

'It seems that now only the chat function works, no active manipulation like with the Claude Chrome extension.'

Read aloud was silent. Whether text-to-speech was switched on and a voice chosen is not recorded.

The 'Current Chat Model Settings' panel lists Keep Alive, Temperature, Seed, Context Window Size (num_ctx), Max Tokens (num_predict), Top K, Top P, Num GPU, Min P, Repeat Penalty, Repeat Last N, TFS-Z, Num Keep, Num Thread, useMmap, useMlock and Reasoning Effort. In the text pasted by the tester only Keep Alive and Seed have a description line. Tooltips were not checked.

'Great model settings, but no explanations so most will miss what all this is.'

The 'Add MCP server' button in the chat does nothing visible; it expects that a server was added in settings first. The tester calls this a UI issue.

'I will deinstall now all of this because it's not relevant. I don't need another search tool, search engine, or another interface to work with local LLMs. I have many of them.'

He wanted a sidebar to the pages he surfs, where he asks questions and gets answers from a local model, sometimes from the internet, and that clicks through tasks like Claude's sidebar. 'This is not a sidebar. This is a new browser.'

'For now, stay with Claude Chrome extension. There I can do all.'