Clay MCP server for your AI

One personal Clay API key. Create it in Clay under Settings, then Account, then API keys, and copy it straight away because Clay shows it only once. It is the only credential Clay's Public API accepts, and it works in the workspace you created it in, so the connection is locked to that workspace. PopMCP checks it by reading the key's user and workspace, stores it encrypted, and sends it from the server as the clay-api-key header.

Yes. Running a function or workflow spends Clay data credits, and on newer Clay plans action executions too; a batch run counts one run per row. Searching Clay's database spends no credits but counts toward your plan's result limits, and Free and Trial plans cap a search at 50 results. get_credit_balance reads both balances without spending anything, and PopMCP's Credits running low trigger fires when either drops below the level you set.

You tell it. Clay's Public API has no endpoint that lists functions or tables, so give the AI the function's Clay URL or its t_ id (workflows use wf_ ids). PopMCP accepts function:t_..., a bare id or the full URL and adds the prefix itself. The function needs API access turned on in Clay under Functions, Details, API, otherwise runs return 404, and the input names come from the function's inputs in Clay.

Yes, up to 20,000 rows per call. run_routine_batch builds the JSONL input file, uploads it to Clay's presigned URL and starts the batch run, a file upload a model cannot do by itself. get_batch_run_rows then checks the run and, once it is complete, downloads Clay's short-lived result file and returns the rows up to 1,000 at a time. For 1 to 100 rows, run_routine_and_wait starts the run, polls it and returns every row's result in one call.

People and companies, queried in Clay's own search language. search_database creates the search and returns the first page (25 rows by default, up to 500), and get_search_results returns the next page while has_more is true. Queries can match people by current job title, company industry and headcount, or companies by the people who work there and by technology such as Salesforce. get_search_query_reference returns every searchable field.

Yes, on Clay's Enterprise plan. query_tables runs a structured query with select, filter, order, group and join across one or more tables, and columns can be named by their titles. There is no list-tables endpoint, so give the AI the table URL or the t_ id from it. Results come back in pages with a cursor, and a record updated mid-scan can appear twice, so ask the AI to deduplicate by id when counts matter.

No. PopMCP hosts it and gives you a Streamable HTTP MCP URL. There is nothing to deploy, no callback URL to register, and no infrastructure to keep running or patch.

Any client that supports a remote Streamable HTTP MCP server: Claude Desktop, Claude Web, ChatGPT custom connectors, Codex and the Codex CLI, OpenCode and Z.ai's ZCode, as well as editors such as Cursor, Windsurf and Zed. Your client signs in to PopMCP with OAuth; the provider credential never leaves the server.

Loading 17 tool definitions into every conversation burns context before the work starts. PopMCP loads a curated 15-tool set covering the everyday workflows and keeps the rest searchable, so your AI pulls a specialist tool only when a task needs it. Every tool stays searchable, and on Scale the owner can load the full catalog for a teammate's connection.

A tool is something your AI calls, and there are 17 here. A trigger is something PopMCP watches for you: 1 Clay events that can start work on the PopMCP side. Triggers never enter your AI client's context, and none of them ask you to paste a callback URL into Clay.