MCP Quickstart

PromptLoop ships a remote Model Context Protocol server so any MCP-capable AI client can run your research tasks, launch batches, search for companies, and check usage — all authenticated with OAuth. There is nothing to self-host.

Server URL

https://promptloop.com/mcp

The transport is Streamable HTTP and auth is OAuth 2.1 (PKCE + Dynamic Client Registration). The first time a client connects it opens a browser window where you sign in and approve access. The connection is scoped to your team and can be revoked anytime from API settings.

The fastest way to connect is the in-app installer: promptloop.com/mcp/install has one-click buttons and copy-paste configuration for every client.

Connect your client

Cursor
Claude
ChatGPT
Codex
Claude Code

Add this to ~/.cursor/mcp.json:

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

Cursor shows a Needs login prompt — click it to authorize with OAuth.

In Claude (web or desktop): Settings → Connectors → Add custom connector, set the URL to https://promptloop.com/mcp, then click Connect and sign in.Custom connectors require a paid Claude plan (Pro, Max, Team, or Enterprise).

Enable Settings → Connectors → Advanced settings → Developer mode. In the Connectors tab click Create, set the MCP server URL to https://promptloop.com/mcp, set Authentication to OAuth, then create and authorize.

codex mcp add promptloop --url https://promptloop.com/mcp

Restart Codex — a browser window opens for OAuth on the first tool call.

claude mcp add --transport http promptloop https://promptloop.com/mcp

Run /mcp inside Claude Code and pick promptloop to authenticate.

Available tools

Tool	What it does
`list_tasks`, `get_task`	Discover saved tasks and their inputs/outputs
`run_task`	Run a saved task on one row; inputs are the fields for that row
`run_dynamic_research`	Ad-hoc research on a single website, no saved task
`create_task`, `edit_task`, `test_task`	Build and refine tasks from natural language
`create_batch`, `list_batches`, `get_batch`, `get_batch_results`	Run and track large async jobs
`cancel_batch`, `retry_batch`, `set_batch_webhook`	Manage batches
`get_company_research`	Retrieve stored structured data for up to 100 companies
`search_companies`	Find companies matching an ICP and export a CSV
`get_usage`, `get_limits`	Check credit usage and plan limits

HubSpot utilities for admins

The following tools are advertised only when the person who authorized the MCP OAuth connection is currently a PromptLoop admin with user mode disabled in the connected team, and that team has HubSpot connected. This is the PromptLoop admin role, not a HubSpot portal role. Ordinary team API keys do not grant this access. Authorization is checked again on every tool call, and HubSpot’s own connection permissions still apply.

Tool	What it does
`get_hubspot_properties`	Inspect company property names, types, enum options, and writability
`get_hubspot_coverage`	Identify dataset domains that are present or missing in HubSpot
`preview_hubspot_sync`	Preview company matching and create/update/conflict/skip decisions without CRM writes
`sync_dataset_to_hubspot`	Sync dataset columns to existing writable HubSpot company properties

Dataset tools take dataset_id, file_id, and domain_column. file_id is the saved CSV’s fileId, not the dataset version’s datasetFileId. Both the dataset and file must belong to the connected team, and the file must be linked to that dataset. Inspect properties and preview matching before syncing. The preview shows domain matching and row decisions, not exact property changes. Sync defaults to update-only and filling empty properties, skips ambiguous domain matches, and never creates HubSpot properties. Set overwrite_existing: true only when replacing existing values is part of the user’s authorized scope. A timed-out sync may still be running; do not automatically retry it.

Setup prompt

Paste this into your assistant so it knows how to drive PromptLoop:

Prompt template

You have access to the PromptLoop MCP server (https://promptloop.com/mcp).

- Use list_tasks to find saved research tasks before creating a new one.
- Prefer one reusable task with multiple related questions/output columns, not one task per datapoint or company. For example, research company description, customer industries, and pricing together.
- Include all related questions, output columns, and formatting requirements in create_task.description, then validate with test_task on a representative input.
- Use run_task for one row of a saved task; its inputs array contains that row's fields in the order returned by get_task, not a list of separate companies.
- Use run_dynamic_research for a one-off website and put all related questions in one queries array.
- For multiple companies needing the same research, prefer create_batch with one row per company, even for modest lists. Reuse the same task; split lists over 10,000 rows into multiple batches.
- Track the returned batch_id with get_batch, then get_batch_results when complete; do not resubmit the same work when checking progress.
- Use get_company_research to retrieve stored structured data before starting new research. Include up to 100 company URLs/domains in one call. It does not browse the web, refresh data, or load generated markdown.
- Use search_companies to build a target list, and get_usage / get_limits to check credits.

Confirm the scope of large/expensive runs with me unless I have already authorized it.

Revoking access

Every connection is backed by a short-lived, auto-rotating team API key. To disconnect a client, revoke its key from API settings — the client will stop working immediately.

Get Started

API Introduction

Server URL

Connect your client

Available tools

HubSpot utilities for admins

Setup prompt

Revoking access

​Server URL

​Connect your client

​Available tools

​HubSpot utilities for admins

​Setup prompt

​Revoking access

Server URL

Connect your client

Available tools

HubSpot utilities for admins

Setup prompt

Revoking access