DeepSieve MCP server

Two ways to connect. This package is the stdio + API-key server — ideal for CI, containers, and services. Where a deployment has the remote server with browser login enabled, that route is easier for a developer at their editor — no key to manage. It is not enabled on every deployment, so see https://deepsieve.ai/developers/mcp for the current endpoint rather than assuming one.

Gives any MCP client (Claude Code, Cursor, Claude Desktop, …) nine outcome-shaped tools over the DeepSieve /v1 API: verify the account, discover the Blueprint schema, start/poll/cancel research runs, read the cited dataset incrementally, export it, and inspect webhooks.

It is a plain API client — it holds no privileged access, only the scoped key you give it (mint one with the Agent preset at /settings/api-keys; it can run research and read results but can't touch your schema or billing).

Install

🛑 0.1.0 does not start, and 0.1.1 fixes it. 0.1.0 declared an unbounded mcp>=1.2.0; mcp 2.x renamed FastMCP to MCPServer, so on 0.1.0 uvx deepsieve-mcp fails with ModuleNotFoundError: No module named 'mcp.server.fastmcp'. Nothing in this package changed — the dependency released a major, months after we published.

0.1.1 caps the dependency at mcp<2 and starts normally, so plain uvx deepsieve-mcp is correct again. If you are pinned to 0.1.0, add one flag: uvx --with 'mcp<2' deepsieve-mcp (or "args": ["--with", "mcp<2", "deepsieve-mcp"] in the JSON below).

The port to mcp 2.x — which is what will let the cap come off — is tracked in #796.

# Claude Code
claude mcp add deepsieve \
  --env DEEPSIEVE_API_KEY=ds_live_... \
  --env DEEPSIEVE_API_URL=https://deepsieve.ai \
  -- uvx deepsieve-mcp

Manual mcpServers JSON (Cursor, Claude Desktop):

{
  "mcpServers": {
    "deepsieve": {
      "command": "uvx",
      "args": ["deepsieve-mcp"],
      "env": {
        "DEEPSIEVE_API_KEY": "ds_live_...",
        "DEEPSIEVE_API_URL": "https://deepsieve.ai"
      }
    }
  }
}

DEEPSIEVE_API_URL defaults to http://localhost:8200 (local dev stack).

Tools

• Orientation: get_account · list_blueprints · get_blueprint · list_members
• Research: estimate_research · start_research · get_run_status · list_runs · cancel_run
• Data: query_entities · export_dataset · get_column_changes · get_cell_changes · list_changes · undo_change · list_held_cells · apply_cell_finding · release_cell · preview_merge · merge_records
• Reports: list_reports · get_report_columns · create_report · get_report_data
• Chat: ask_research_agent · get_chat_thread · confirm_chat_action · dismiss_chat_action · revert_chat_action
• Blueprints: propose_blueprint_change · create_blueprint · get_blueprint_draft · answer_blueprint_question · adjust_blueprint · approve_blueprint
• Webhooks: list_webhooks

A new draft stops at ready or at awaiting_input. awaiting_input means it has follow-up questions, and only answer_blueprint_question (or approving as it stands) moves it on.

Full API docs: /developers on your deployment (also /llms.txt). PyPI/npm publishing and a hosted Streamable-HTTP OAuth server are tracked follow-ups (issue #116 Phase-2).