Connect PromptLoop to your AI tools
One MCP server for Claude, ChatGPT, Codex, Cursor, and Claude Code. Sign in once with OAuth, no API keys to paste.
Client
Install in one click
Add to Claude- 1.Open Settings → Connectors → Add custom connector.
- 2.Set the URL to https://www.promptloop.com/mcp.
- 3.Click Connect and sign in with OAuth when prompted.
- 4.Custom connectors require a paid Claude plan (Pro / Max / Team / Enterprise).
Or add this connector URL
https://www.promptloop.com/mcpOr copy a setup prompt for Claude
Help me connect the PromptLoop MCP server in Claude.
PromptLoop is a remote MCP server (Streamable HTTP + OAuth 2.1) that runs AI web research to enrich B2B data. Server URL: https://www.promptloop.com/mcp
Install steps:
1. Open Settings → Connectors → Add custom connector.
2. Set the URL to https://www.promptloop.com/mcp.
3. Click Connect and sign in with OAuth when prompted.
4. Custom connectors require a paid Claude plan (Pro / Max / Team / Enterprise).
Auth is OAuth 2.1 — on first connect a browser window opens to sign in to PromptLoop and approve access. It is scoped to my team; there are no API keys to paste.
Once connected, confirm these tools are available and tell me: list_tasks, get_task, run_task, run_dynamic_research, create_task, edit_task, test_task, create_batch, list_batches, get_batch, get_batch_results, cancel_batch, retry_batch, set_batch_webhook, get_company_research, search_companies, find_work_emails, get_work_email_results, retry_work_emails, get_usage, get_limits.
Usage guidance:
- Use list_tasks / get_task to find and inspect a suitable saved task before creating one.
- Prefer one reusable task with multiple related research items (questions/output columns) for the same workflow. For example, extract company description, customer industries, and pricing together instead of creating a separate task for each datapoint. This reduces repeated task setup and research runs. Split tasks when they require different inputs or independent workflows.
- Include all requested related questions, output columns, and formatting requirements in create_task.description. Test the task on a representative input with test_task before scaling it. For one-off research on one website, include all related questions in a single run_dynamic_research.queries array.
- A task defines the research workflow; a batch applies that task to many items (companies/websites). Reuse the same task across companies instead of creating one task per company.
- run_task executes ONE row synchronously. Its inputs array contains that row's input fields in the order returned by get_task, usually a single company website; it is not a list of companies to process independently.
- For multiple companies needing the same research, prefer create_batch with one row per company, even for a modest list, instead of repeated run_task calls. Support up to 100,000 rows per batch. For large files (including 50,000 rows), use prepare_batch_upload, upload the JSONL file directly with an HTTP client using the returned URL and upload_fields, with the file as the final multipart field, then call create_batch_from_upload. Do not split the file into small batches or paste its contents into chat. Use run_task or run_dynamic_research for a single company, and create_batch for work that may exceed the synchronous timeout.
- Track a batch with get_batch, then fetch get_batch_results when complete (download=true for large outputs). Reuse the returned batch_id for follow-up checks instead of submitting the same work again.
- Use get_company_research to retrieve stored structured data before running new research when the user asks about known companies. Put up to 100 URLs/domains in one call instead of making one call per company. Treat missing data, team research marked stale, or records marked enrich_failed, needs_refresh, or failed_refresh as a reason to use a task; this lookup does not browse the web or refresh data.
- To find work emails for people, use find_work_emails rather than building a research task. For a file or spreadsheet, pass the whole thing in one call with csv_url (direct link or Google Sheets/Drive share link), file_id (a PromptLoop dataset), or csv_text, up to 10,000 rows; do not split it into batches or rebuild it as people arrays. Headers are detected automatically and the mapping used is returned; pass columns only to correct it. Each person needs a name and a company website/domain, LinkedIn profile URL, or work email; company name alone is not enough, so find the website first when it is missing.
- Run find_work_emails with dry_run: true first on any file to see row counts, rejected rows with reasons, duplicates, the column mapping, and the credit ceiling (5 credits per email found, nothing for people not found), then confirm the spend with the user and launch with the same input.
- Poll get_work_email_results with the returned job_id instead of resubmitting the file; when status is COMPLETED, call it with format: "csv" to get a download link with the original columns plus email, verified, confidence, email_status, source and status. Only report an email as verified when verified is true; confidence "probable" means a likely but unconfirmed address. If rows failed, retry_work_emails reruns them without double charging.
- If a bug, unavailable capability, or missing data prevents fulfilling the user goal after trying relevant tools, call report_issue with a concise description and the tools tried. Exclude secrets and raw customer files. Reuse the request_id when retrying notification delivery.
- Check get_usage / get_limits before large or expensive runs. Confirm the scope with the user unless they have already authorized it.
If this MCP connection was authorized by a current PromptLoop admin and the same team has HubSpot connected, additional tools are available: get_hubspot_properties, get_hubspot_coverage, preview_hubspot_sync, sync_dataset_to_hubspot. Inspect properties and preview matching before a user-authorized sync. These tools are hidden from other connections.