Connect your AI assistant to Pave Agent

What is this?

If you already work in an AI assistant such as Claude or ChatGPT, you can ask Pave Agent compensation questions without leaving it. Your assistant sends the question to Pave Agent, Pave Agent runs the analysis in Pave with your data and your access, and the answer comes back to your assistant.

The connection uses the Model Context Protocol (MCP), a standard way for AI assistants to connect to other tools. You add Pave Agent to your assistant once, log in to Pave, and then ask questions in plain language.

These assistants are supported:

• Claude, in the web and desktop apps
• Claude Code
• Cursor
• ChatGPT

Before you connect

You need:

• Pave Agent and the Pave Agent connector turned on for your company. If you do not see Pave Agent in Pave, or are not sure whether the connector is turned on, ask your Pave representative.
• The company admin role in Pave. Only company admins can use Pave Agent through an AI assistant.

Access is checked when you ask a question, not when you connect. If you do not have access, you can still add the connector and log in, but your first question returns a "not authorized" message.

You also need the Pave Agent connector address:

https://mcp.pave.com/mcp

Connect your assistant

In Claude, you add Pave Agent from Claude's connector directory. In the other assistants, you add it as a custom connector. Either way, you then log in to Pave the way you normally do, for example with your company's single sign-on.

Menu names in these assistants change from time to time. If you cannot find a step, look for the option to add a custom or remote MCP server.

Claude

Pave Agent is listed in Claude's connector directory. These steps work in the Claude web and desktop apps. [UNVERIFIED: confirm the directory button labels, menu path, and desktop app availability against the live listing]

1. Open the Pave Agent listing in Claude's connector directory. You can also find it in Claude: open Settings, then Connectors, choose Browse connectors, and search for Pave Agent.
2. Choose Connect and log in to Pave.

Claude Code

1. Run this command in your terminal:

   claude mcp add --transport http -s user pave https://mcp.pave.com/mcp
   

2. Start Claude Code and type /mcp.

3. Choose pave, then Authenticate, and log in to Pave.

Cursor

1. Open Cursor Settings, then MCP, and add a new MCP server.

2. Use the URL https://mcp.pave.com/mcp. If you would rather edit the configuration file directly, add:

   {
     "mcpServers": {
       "pave": { "url": "https://mcp.pave.com/mcp" }
     }
   }
   

3. When Cursor prompts you, log in to Pave.

ChatGPT

1. In ChatGPT settings, open the connectors or apps section. Turn on developer mode if your plan requires it for custom connectors.
2. Create a custom connector with the URL https://mcp.pave.com/mcp and choose OAuth for authentication. Leave the client ID and secret empty.
3. Connect and log in to Pave.
4. In a chat, turn on the Pave Agent connector from the tools menu.

Ask a question

Ask in plain language. Your assistant passes the question to Pave Agent.

"What is the median base salary for Software Engineer II in San Francisco?"

Here is what happens:

1. Your assistant starts the analysis. Pave Agent begins working right away, and your assistant does not have to wait on a single long request.
2. Pave Agent does the work in Pave. A focused question can take seconds. A complex analysis can take several minutes.
3. Your assistant checks in and shows you what Pave Agent is working on until the answer is ready.
4. You get the answer in your assistant.

Pave Agent answers as it would in Pave. It uses your access, your company's data, and your Guidance and Skills.

Tips for better answers

• Ask one focused question at a time.
• Include the specifics that matter, such as job family, level, location, and time period.
• Ask follow-ups in the same chat. Your assistant can send a follow-up to the same Pave Agent conversation, so Pave Agent builds on its earlier questions and answers there. Wait for the current answer to finish before asking the next follow-up. To start fresh, ask your assistant for a new Pave Agent analysis.
• Put the details in your question. Pave Agent sees the questions your assistant sends it, not the rest of your assistant's chat. If you shared notes or files with your assistant, include the details Pave Agent needs in your question.
• Pick up in Pave whenever you like. Every conversation, including its follow-ups, also appears in your Pave Agent history in Pave. Open it there to keep going with the full context.

Things to know

Answers come back as text. Your assistant receives Pave Agent's written answer. Source chips, files Pave Agent creates such as dashboards or presentations, and CSV downloads are not included. To get those, ask the question in Pave Agent in Pave, or open the conversation from your history there.

Long analyses have a time limit. An analysis that runs longer than 20 minutes stops and returns an error.

Your assistant should not resend the same question while waiting. If it does, the analysis runs twice. If your assistant seems stuck, wait for it to report a result or an error before asking again.

Pave Agent does not change your compensation data, whether you ask in Pave or through your assistant. See the article on getting started with Pave Agent.

Troubleshooting

"You are not authorized to use Pave Agent, including through this MCP connector.": You do not have the company admin role in Pave, or Pave Agent or the Pave Agent connector is not turned on for your company. Retrying will not help. Ask your Pave admin or your Pave representative for access.

Logging in fails with "no connections enabled": Your company's login method is not yet enabled for the Pave Agent connector. Contact your Pave representative.

Logging in fails with "Invalid or expired OAuth state.": The login took more than 10 minutes to finish, or the connector was set up incorrectly. Remove the connector from your assistant, add it again using the steps above, and complete the login right away.

"The analysis failed.": The analysis did not finish, for example because it reached the 20-minute time limit. Ask a narrower question rather than resending the same one.

"The analysis failed: no progress reported for 3 minutes." or "The analysis failed: no result within 22 minutes.": The analysis stopped before it finished. Ask the question again.

"The analysis could not be started within 30 seconds.": Pave Agent did not receive the question in time. Asking again starts a new analysis from the beginning.

The connector shows as disconnected: Your login to Pave has expired. Reconnect from your assistant's connector settings and log in again.