Skip to content

Solve problems with the AI agent

What to do when the sign-in, the access, a project or a check does not work.

Last updated: November 20, 2026


Most problems come from the chosen access, the project, your plan or your agent’s settings. Here is what you can do in each case.

Connecting and signing in

The sign-in window does not open

Start the sign-in again in your agent. In Claude, open sitebrunch under “Customize” and “Connectors” and click “Connect” or “Reconnect”. In Claude Code, run /mcp and pick sitebrunch. If your browser blocks pop-ups, allow them for sitebrunch.

The agent reports “The connection was ended in sitebrunch”

Someone clicked “Disconnect all agents” in the account settings. Connect your agent again: in Claude, click “Reconnect” next to sitebrunch; in other agents, start the sign-in again.

Claude reports “MCP server connection lost”

The connection between Claude and sitebrunch was briefly interrupted, or a query took too long. Ask your question again. If it keeps happening, start a new chat. Continue a running audit there with “Continue the sitebrunch audit”.

Access, projects and plan

The agent cannot change anything

Your connection only has “Read only”. The agent then sees no tools that change data. Disconnect sitebrunch in your agent, connect it again and choose “Read and write”.

The agent says no project is selected

Tell your agent which project to work with, for example: “Work with the sitebrunch project for my website.” It then selects the project itself.

A project is missing

The agent sees the projects of every workspace you are a member of. If a project is missing, ask the workspace owner to invite you.

Scans, PDFs, analytics or notes are not available

These tools need a workspace on the Starter plan or higher. Analytics and conversions also need a project whose plan includes them. Without these plans, your agent can read the results of the automated check and work on comments.

In Claude

Claude keeps asking for permission

“Always allow” only ever applies to one tool. Open sitebrunch under “Customize” and “Connectors” and set the tools to “Always allow” under “Tool permissions”. On Claude Team and Enterprise, an admin first has to turn on “Allow ‘Always allow’ for connector tools” in the organization settings.

sitebrunch is missing in the Claude phone app

You cannot add a connector on your phone. Connect sitebrunch once in the browser or the desktop app. It then shows up on your phone too.

In ChatGPT

You cannot find “Add custom MCP server”

In a ChatGPT Business or Enterprise workspace, admins decide who may add custom MCP servers. Ask your admin to allow it or to add sitebrunch. In the desktop app you can also add sitebrunch under “Settings” and “MCP servers”.

New tools are missing or calls fail

In a ChatGPT Business or Enterprise workspace, ChatGPT uses the tools as they were when the server was published. Ask an admin to refresh the sitebrunch tools in the workspace settings. In the desktop app, click “Restart” under “Settings” and “MCP servers”.

ChatGPT does not use sitebrunch

Mention sitebrunch with @ in your message. The selection only applies to that one message.

When testing with an AI agent

Testing with an AI agent is still at an early stage and only enabled for workspaces that test it. The setup there solves most problems: the right agent, agent-browser, the start prompt and the permissions.

The agent says guided audits are not enabled for this workspace

When you continue an audit with its audit ID, your agent selects the right project itself. For a new audit, check that it selected the right project, and otherwise tell it which website is being tested.

If the message also appears in the right project, the feature is not enabled for your workspace yet. Write to us at support@sitebrunch.com.

The agent says the website cannot be reached or the test browser is missing

Before the first step, your agent checks whether its browser can load the pages. If it reports blocked domains, allow exactly the domains it names. How that works in your agent is described under When your agent asks you to allow domains. Then start a new chat and write “Continue the audit”.

If the browser is missing, install agent-browser with npm install -g agent-browser && agent-browser install. Or give your agent another browser such as Chrome DevTools MCP.

The agent is not allowed to install agent-browser

Some agents refuse installs on their own when they approve commands automatically. Allow the install when your agent asks, or run the command yourself in your terminal:

npm install -g agent-browser && agent-browser install

Without agent-browser, your agent tests with another browser, for example Chrome DevTools MCP. That takes longer, and you should not use the computer meanwhile.

The agent stops after the automated scan, and the audit shows 0 completed steps

Your agent did not start subagents. Without them it tests a little by itself and then stops. This is how you solve it:

  • The normal ChatGPT chat cannot do it. Switch to “Work” mode or to Codex.

  • In “Work” and Codex, ChatGPT only starts subagents when you ask. Use the start prompt or write: “Run the audit with subagents.”

  • An audit runs most reliably in Claude Code.

The agent says it cannot start subagents

sitebrunch checks this before the start and stops instead of testing halfway. In “Work” or Codex, write: “Run the audit with subagents.” In the normal ChatGPT chat, switch to “Work” or Codex first.

A tester reports “Cannot start subagents.”

Testers need subagents of their own that judge the screenshots. In Codex this is off at first. Add the line max_depth = 2 under [agents] in ~/.codex/config.toml and restart Codex. In Claude Code it works on its own, unless CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH is set to 1. More under Allow subagents inside subagents. Then continue the audit in a new chat.

The agent asks for permission or confirmation at every step

Allow the sitebrunch tools before you start, as described under Give permissions once. If the agent asks between steps whether to continue, use the start prompt. It tells the agent to keep working without pausing.

agent-browser does not start in Codex or the ChatGPT desktop app

The Codex sandbox blocks agent-browser, and Chrome crashes at start. Allow the command to run outside the sandbox when Codex asks. Or let your agent test with the browser built into ChatGPT.

The agent tests in a window, and the keyboard results are wrong

A visible browser window loses focus as soon as you click into another app. Focus outlines then disappear, and the keyboard checks go wrong. Do not use the computer during the audit, or use agent-browser. It runs without a window.

The dashboard says “No action for 20 min.”

The dashboard only counts calls to the sitebrunch tools. A tester that walks through a page with the keyboard for a long time only calls them when it saves a finding. Look at the chat: if no question is waiting for you there, the tester is still working.

Screenshots hang when the display turns off

When a Mac display goes to sleep, screenshots in the browser hang. Keep this command running in a terminal during the audit. It keeps the display awake until you stop it with Ctrl+C:

caffeinate -d

The chat breaks off in the middle of the audit

The audit stays open in sitebrunch. Start a new chat and write: “Continue the sitebrunch audit <audit ID>.” The audit ID is in the dashboard, in the address of the audit. Your agent picks up at the next open pass. It marks steps from the old chat for your review, because it does not know their reports. Have them retested if needed.

Requests from the audit arrived on your website

The testers send forms to check error messages. This creates real entries, for example contact requests or sign-ups. For the next audit, write in the prompt which forms your agent must not send.

The agent follows old instructions

A chat keeps the instructions it loaded at the start. Start a new chat after a sitebrunch update.

Other agents

GitHub Copilot shows no tools

In GitHub Copilot the tools only work in agent mode. Switch the chat to agent mode.

Info
If you are stuck, write to support@sitebrunch.com. Tell us your agent and the error message it shows.