MCP endpoint

Give agents a trust-record checkpoint where they work.

MCP clients should read get_trust_record before public browser work. Existing preflight users keep working through the alias path.

Discovery JSON Reporting API

Integration path

Agent ready

Read

public remote mode

HTTP

streamable JSON mode

2025-06-18

protocol

npx

stdio package

Checkpoint framing

The one call before your agent acts

Registry kits and connector listings should describe CrawlDex as the tool that returns the Agent Trust Record before a website task starts. That record carries the verdict, confidence, blockers, evidence, publisher claim state, and dispute path.

Copyable local snippets set `CRAWLDEX_CHANNEL` so weekly funnel reports can attribute installs without changing canonical CrawlDex URLs.

crawldex-mcp sends a random anonymous instance ID with API calls so CrawlDex can count distinct agents. It contains no personal data, is stored server-side only as a salted hash, and is never linked to your identity or IP. Disable with `CRAWLDEX_NO_INSTANCE_ID=1` -- functionality is identical without it.

Safety model

Read tools here, writes through the API

1Trust recordCall get_trust_record before public browser work and keep the issued record id with the run.

2Preflight aliasExisting preflight users keep the same compatibility language while migrating to get_trust_record.

3ReportAfter the attempt, submit a redacted report through the API so writes stay explicit and auditable.

Available tools

What an agent can ask

tool get_trust_record

Call before an agent attempts a public website task. Returns the citable Agent Trust Record payload.

tool preflight

Alias of get_trust_record for existing users; nothing breaks while clients move to the trust-record name.

tool resolve_intent

Call first when the user gives a plain-language website task. Returns canonical task candidates before the trust-record read.

tool describe_task

Inspect success criteria, posture, related tasks, and known blockers before choosing a route.

tool get_recipe

Fetch available public recipe metadata after preflight says a recipe exists.

tool get_alternatives

Find better sites for the same task when the selected route is blocked, stale, or low confidence.

tool explain_blockers

Explain observed blockers so the agent can tell whether to continue, pause, or ask for user presence.

tool list_tasks

Discover task keys by category for agents that need to map their own internal task labels.

api POST /api/v1/runs

Use the reporting API for outcome writes; public remote MCP write tools are disabled.

Install status

Channel-tagged MCP snippets

Claude Code

bash

claude mcp add crawldex -- env CRAWLDEX_API_ORIGIN=https://crawldex.com CRAWLDEX_CHANNEL=pack-claude-code npx -y crawldex-mcp

✓ verified working 2026-07-22 · https://crawldex.com · mcp-claude-code-command · 2026-07-22T21:21:21.001Z

Claude Desktop

json

{
  "mcpServers": {
    "crawldex": {
      "command": "npx",
      "args": [
        "-y",
        "crawldex-mcp"
      ],
      "env": {
        "CRAWLDEX_API_ORIGIN": "https://crawldex.com",
        "CRAWLDEX_CHANNEL": "pack-claude-desktop"
      }
    }
  }
}

✓ verified working 2026-07-22 · https://crawldex.com · mcp-claude-desktop-config · 2026-07-22T21:21:21.002Z

Cursor MCP JSON

json

{
  "mcpServers": {
    "crawldex": {
      "command": "npx",
      "args": [
        "-y",
        "crawldex-mcp"
      ],
      "env": {
        "CRAWLDEX_API_ORIGIN": "https://crawldex.com",
        "CRAWLDEX_CHANNEL": "pack-cursor"
      }
    }
  }
}

✓ verified working 2026-07-22 · https://crawldex.com · mcp-cursor-config · 2026-07-22T21:21:21.002Z

VS Code MCP JSON

json

{
  "servers": {
    "crawldex": {
      "command": "npx",
      "args": [
        "-y",
        "crawldex-mcp"
      ],
      "env": {
        "CRAWLDEX_API_ORIGIN": "https://crawldex.com",
        "CRAWLDEX_CHANNEL": "pack-vscode"
      }
    }
  }
}

✓ verified working 2026-07-22 · https://crawldex.com · mcp-vscode-config · 2026-07-22T21:21:21.002Z

Cursor deeplink

text

cursor://anysphere.cursor-deeplink/mcp/install?name=CrawlDex&config=eyJjb21tYW5kIjoibnB4IiwiYXJncyI6WyIteSIsImNyYXdsZGV4LW1jcCJdLCJlbnYiOnsiQ1JBV0xERVhfQVBJX09SSUdJTiI6Imh0dHBzOi8vY3Jhd2xkZXguY29tIiwiQ1JBV0xERVhfQ0hBTk5FTCI6InBhY2stY3Vyc29yIn19

✓ verified working 2026-07-22 · https://crawldex.com · mcp-cursor-deeplink · 2026-07-22T21:21:21.002Z

GPT Actions

json

{
  "openapi_url": "https://crawldex.com/openapi.json",
  "authentication": "none",
  "headers": {
    "x-crawldex-channel": "pack-gpt-actions"
  },
  "channel": "pack-gpt-actions",
  "when_to_call": "Before attempting a public website task, especially checkout, cancellation, refund, booking, support, forms, account, login, or policy routes.",
  "privacy_boundary": "Send only public site, task, and agent-stack metadata. Do not send secrets, cookies, MFA codes, payment data, private account data, screenshots, raw traces, or user-specific identifiers."
}

✓ verified working 2026-07-22 · https://crawldex.com · mcp-gpt-actions-config · 2026-07-22T21:21:21.002Z

Remote MCP

json

{
  "mcpServers": {
    "crawldex": {
      "type": "http",
      "url": "https://crawldex.com/mcp",
      "headers": {
        "x-crawldex-channel": "web-developers"
      }
    }
  }
}

✓ verified working 2026-07-22 · https://crawldex.com · mcp-remote-config · 2026-07-22T21:21:21.002Z

Client kits

Registry copy uses the checkpoint claim

Claude Code, Claude Desktop, Cursor, VS Code, GPT Actions, and other MCP-capable clients should list CrawlDex as a read-only checkpoint for public website tasks: fetch the trust record first, then continue, hand off, or avoid the route.

Read API guide

Outcome writes

Reports stay outside public MCP

Public remote MCP write tools remain disabled. Use the reporting API for explicit redacted outcome writes, and keep credentials in HTTP headers rather than connector sessions.