Do I give AI my password or cookies?
No. Requests run in your signed-in browser, and login credentials and website CSRF values are not exported as credential fields in tool results. Requested account details, private content, or other results may still contain personal information and are returned to your AI client.
Do I still need the extension after installing an MCPB?
Yes. MCPB installs the local MCP server, while the extension executes operations in website pages. The two components connect automatically on the same computer without pairing codes, port entry, or website developer keys.
How many tokens does it save?
There is no fixed saving guarantee. Only three tools appear initially, operation schemas are discovered for the relevant site and task, and lists use selected fields and length limits. Actual token use depends on the model, content, and workflow. maxChars is a character budget.
Can I use a cloud MCP server or remote browser?
The current connection requires MCP and the browser extension on the same computer. The bridge listens only on loopback. The project does not provide a public remote-browser proxy and does not require exposing a port to the internet.
Will every write ask for another confirmation?
The extension adds no second confirmation. AI should only perform writes requested by the user. Ask for a draft first if you want to review content before explicitly requesting publication. Your AI client may have separate tool-approval controls.
Can a comment still be sent after cancellation?
Yes. Cancellation can stop waiting or prevent steps that have not run; it cannot undo a request already received by the website. The same applies to timeouts and disconnections. Check the final website state before sending the same action again.
Why does an API fail when its page opens normally?
Websites may apply different permissions and verification to different access paths. Some Zhihu reads have normal-page fallbacks; other restrictions are reported as errors. The project does not bypass login, verification, signatures, or access controls.
Why does the next page return PAGE_NOT_READY?
The topic page has not loaded new cards and has not explicitly reached the end. Retry the same read later. If the result is TARGET_PAGE_REQUIRED, AI can call site.context with the provided normal page URL. After navigation, select the target again and use an appropriate cursor.
Can several AI clients connect?
Local clients can share the bridge, with separate target handles and output caches. They still operate the same browser and website accounts, so avoid conflicting instructions against the same content.
Is it free, and is it an official platform integration?
The source is available under the MIT license. Your AI provider, network, and websites may have their own charges or terms. This is an independent project and does not claim official partnership or endorsement by Reddit, Zhihu, OpenAI, or other named platforms.
What about macOS, Linux, or phones?
Windows is the tested environment, with Node.js 22+ and Chrome or Edge 120+ required. macOS and Linux have not been tested. Mobile browsers are outside the current validation scope. Cross-platform source code is not evidence of tested support on those systems.