Clay MCP: connect research to your AI assistant

Clay MCP lets an outside AI assistant search and enrich data and run enabled Clay Functions for a signed-in user. The connection goes from your assistant to Clay; it does not mean Clay can connect to every outside tool. Clay’s connection guide↗ (opens in a new tab) documents that direction and its browser sign-in flow. Keep the tested instructions in fenbs when you want another supported assistant to continue the work without guessing which access checks passed.

What should you check before connecting?

MCP, the Model Context Protocol, is a standard way for an AI app to use an outside tool. Here, Clay supplies the tools and your assistant is the client that calls them. You need both a working connection and a clear answer about what that connection may do.

Ask your Clay workspace admin to confirm your access, spending budget, and intended Functions before testing. Clay documents a Sales Rep permission type for MCP-only access, while admins manage the workspace connection policy. An unrecognized outside client can be blocked by that policy. See Clay’s MCP settings↗ (opens in a new tab) for the controls that apply to your workspace.

• Choose one research question that you can check by hand.
• Use a public company website or another approved test input.
• Write down the actions you expect and the changes you do not want.
• Agree on a small spending budget and who can raise it.

For example, ask for one company’s published support page and a source link. Avoid beginning with an instruction to investigate an entire market. A small question makes unexpected searches, duplicate operations, and missing evidence easier to spot.

How do you set up the connection?

Use the documented Clay endpoint, https://api.clay.com/v3/mcp, in an assistant that supports the required remote connection. Clay’s platform connection documentation↗ (opens in a new tab) describes OAuth with PKCE, a browser sign-in method with an extra check that protects the authorization exchange. It is also the reference for developers registering their own client.

1. Have the workspace admin open MCP in the Clay homepage sidebar and confirm your access and credit budget.
2. For a specific Function you need, have the admin check its integration settings and the MCP for reps option.
3. In your chosen assistant, follow that app’s remote-MCP connection instructions using the Clay endpoint above.
4. Complete the browser authorization as the Clay user whose access should apply.
5. Return to the assistant and inspect the available Clay tools before requesting any research.

The assistant’s connection screen belongs to that assistant, so do not look for a universal Clay button there. If a developer is building the client, use the vendor registration instructions rather than inventing an API-key shortcut. For the wider connection model, read what an AI connector does.

How do you test the tools without running a large job?

Begin by asking the assistant to describe its intended operation before it calls a tool. Your goal is to compare the question, the selected tool, and the returned evidence. A connection that succeeds technically can still be a poor fit for the task you meant to delegate.

Research this public company website: [URL].
Find its published support contact page.
Use only the Clay tools needed for this one question.
Do not run a bulk job or change an outside system.
Return the page URL, the finding, and anything you could not verify.
Before acting, state which tool you intend to use.

Open the returned source yourself. Check that it is the intended company, that the relevant statement is actually present, and that the answer has not turned an inference into a fact. Record what happened if the assistant called more tools than expected. That observation is useful even when the final answer was correct.

Clay says outside calls use the same workspace Actions and Data Credits as equivalent Clay work, and one assistant message can trigger several operations. A zero credit limit blocks further actions. Keep those controls in mind before increasing the test size; the MCP settings guide↗ (opens in a new tab) explains spending and access together.

Why might access or a Function be missing?

Separate a sign-in problem from an access problem. If authorization cannot finish, check the client registration and workspace connection policy with your admin. If sign-in works but an expected Function is absent, check whether that Function was enabled for the intended users. Repeating the research prompt will not settle either question.

Do not assume that successful connection exposes every table. Clay’s documentation limits reach through exposed tools, Functions, and Audiences settings. Treat an unexpected refusal as information about the boundary, then ask the admin to verify the intended permission rather than trying another route around it.

Write down the exact symptom, the operation you attempted, and the non-sensitive part of the result. Leave tokens and personal records out. An explanation such as “sign-in succeeded; the required Function was not available” gives the next person a useful starting point. The guide to system prompts and MCP explains why instructions and connected tools have different roles.

What should you keep after a successful test?

Keep a short recipe: the question, approved inputs, required Function, expected output, source-check method, and stopping condition. Add the date you checked the workflow and who owns its access. Store the result of an access check, never the credential that made the check possible.

Keep company research separate from the instructions used to obtain it. If your next task is creating a reusable table research agent, follow the Claygent guide. For organizing several approved recipes, see the AI workflow library. The connection itself is only one part of repeatable work.

A brief record in fenbs can tell the next assistant which checks passed and which question remains open. That prevents an old successful sign-in from being mistaken for permanent permission to run any job.

How to keep your Clay connection checks with fenbs

Keep the useful work so you can carry on from exactly where you left off, on any device and in any other AI you connect. Save the tested recipe and the next step, while leaving the source data in its approved home.

1. Create a free account.
2. Add a task for the Clay research you want to repeat.
3. Save the approved question, tool name, source checks, and any access problem you still need help with.
4. Choose a supported assistant from Integrations and connect it to fenbs.
5. Before you stop, save what worked and the next small test so another assistant can carry on.

fenbs does not replace Clay’s admin controls or credit budget. Clay’s checked documentation does not describe a native way to add fenbs inside Clay; use a supported outside assistant for the shared work record. Explore the Connect your AI tools hub for related connection questions.