MCP quickstart — connect CitationBench to Claude Code, Cursor & more

Add CitationBench's hosted MCP server to your client (Claude Code, Cursor, Claude Desktop, Windsurf, ChatGPT Apps, or any custom MCP client). Make your first tool call. Sixty seconds.

Step 0 (optional) — try without signing up

Skip the key — connect in demo mode. Every tool returns shape-complete synthetic data.

claude mcp add citationbench https://mcp.citationbench.com/mcp

Then in Claude Code:

> List the CitationBench skills available.

Claude calls agent.skills.list and renders the catalog. You're in demo mode (every response has "demoMode": true).

This is how AI agents can build a full CitationBench integration before the human even creates an account.

Step 1 — Create an API key

See Authentication.

export CITATIONBENCH_API_KEY=sk_live_***
export WORKSPACE_ID=ws_***

Step 2 — Connect (Claude Code shown by default)

claude mcp add citationbench https://mcp.citationbench.com/mcp \
  --header "Authorization: Bearer $CITATIONBENCH_API_KEY" \
  --header "X-Workspace-Id: $WORKSPACE_ID"

Verify:

claude mcp list

You should see citationbench with status connected.

→ Other clients: Claude Desktop, Cursor, Windsurf, ChatGPT Apps, custom.

Step 3 — Your first tool call

Open Claude Code in any project:

claude

Ask naturally:

> Research keywords for project management software, fast depth, limit 50.

Claude calls research.keyword.research with the right input. You see the request, the response, and the credits charged inline.

Step 4 — Compose tools

The point of MCP is natural composition. Try:

> Research keywords for project management software (50 results, fast). Then for each
  PURCHASE-intent one, run a serp gap analysis. List the EASY_WIN ones.

Claude orchestrates: research.keyword.research → research.keyword.search (filtering by intent) → research.serp_gap.bulk → renders the EASY_WIN subset.

That's four tools, one prompt, no glue code.

Step 5 — Invoke a skill

For multi-step workflows, you invoke a skill (a packaged composition the agent runs):

> Bootstrap acme.com — full SEO and GEO research. Pause at each step for me to approve.

Claude calls agent.invoke with skill: "bootstrap_brand" and the right input. The MCP server streams progress as notifications.

Workspace scoping

You have three options.

Per-connection (recommended for solo workflows)

Set X-Workspace-Id at connection time. Every tool uses that workspace.

claude mcp add citationbench https://mcp.citationbench.com/mcp \
  --header "Authorization: Bearer $KEY" \
  --header "X-Workspace-Id: ws_acme"

Per-call (agency master keys)

Don't set the header. Each tool accepts a workspace parameter:

> Run a rank check on the "beta" workspace.

Claude passes workspace: "ws_beta" to distribute.track_rank.

Bulk (cross-workspace ops)

> Check ranks across every client workspace, alert me on drops of 5+.

Claude uses workspaces.bulk_action.

Approval mode

For destructive tools (produce.publish, link_building.campaign.send_email, indexing.gsc.submit), Claude Code will surface the approval prompt before each call. You can also set workspace-level approval policies in the dashboard so CitationBench itself holds the invocation at WAITING_APPROVAL until a human decides — even if Claude says go.

What you just learned

• One command connects you. Demo mode works without a key.
• Natural-language prompts trigger the right tools.
• Skills (like bootstrap_brand) wrap multi-step compositions.
• Workspace scoping has three modes; pick the one that matches how you work.

Next

• Authentication — keys, scopes, demo mode in detail
• Workspaces & multi-brand
• MCP tools · Catalog — every tool listed
• MCP tools · Setup — Claude Code
• Playbook: Build an SEO agent in Claude Code