Connect any MCP agent

ez-crawl exposes a standard MCP server, so your agent is not tied to Claude, Codex, or a particular model provider. Use any client that supports remote MCP over streamable HTTP and a Bearer authorization header, with a model capable of using tools.

Claude Code and Codex have ready-to-use setup options in the app. They are convenient presets for the same MCP endpoint, not the only supported ways to connect.

Get your connection details

1. Open Settings → AI Agent → Create token. Copy it: it is shown once.
2. Copy the MCP endpoint displayed there. It ends in /api/mcp.
3. Choose a ready-made client option below, or configure your preferred MCP client with that endpoint and Authorization: Bearer <token>.

Claude Code and Codex

Select the corresponding option in Settings → AI Agent, then copy and run its generated command or configuration. Those options fill in the endpoint and authentication for your deployment.

Cursor

Add ez-crawl as a remote server in Cursor’s MCP configuration. For example, in your personal ~/.cursor/mcp.json:

Cursor MCP configuration

{
  "mcpServers": {
    "ez-crawl": {
      "url": "https://app.ezcrawl.ir/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_EZ_CRAWL_TOKEN"
      }
    }
  }
}

Replace the token with your personal access token. The endpoint is the same one shown in ez-crawl Settings. Merge this entry with any existing servers. Keep the token in personal configuration, outside version control. Confirm the server’s tools are available before asking Cursor’s agent to build a workflow. See Cursor’s MCP configuration reference.

Ollama and other models

You can use a local model served by Ollama, or a model from another provider, through an MCP-capable agent or chat application. Configure ez-crawl’s remote endpoint and Bearer header in that application, then choose a model that supports tool calling.

Ollama serves the model; the agent application manages the MCP connection, executes tool calls, and returns their results to the model. A model by itself does not connect to an MCP server. See Ollama’s tool-calling documentation.

The same approach applies to other MCP-enabled editors, chat applications, and custom agents, whether their model is Claude, GPT, Gemini, Qwen, or another tool-capable model. Check the client’s transport and authentication support rather than looking for a model-specific ez-crawl integration. Local-stdio-only clients need a compatible remote-MCP bridge; they cannot use the hosted URL directly.

Verify the connection

Ask your agent to call building_guide() and list the available node types. Then try a small request:

Build an ez-crawl workflow that reads https://example.com, answers “What is this page for?”, and displays the answer. Use depth zero and one page. Validate before running, inspect the output, and give me the workflow id.

Open the saved workflow in ez-crawl to inspect the same graph your agent created. Follow your first workflow for the complete verification steps.

What the agent does

It studies the site with its own tools (such as Playwright), builds the workflow node by node, validates and runs it, reads each node’s output, and fixes what is wrong. Open the workflow in the app and watch it appear.

While it works

• The canvas is locked: no moving, connecting, deleting, saving or running.
• If you have unsaved edits, nothing changes under you; a banner offers to load the agent’s changes or keep yours.
• When it starts a run, the run panel opens and shows progress live.

Safety

Caution

A token acts as you, with full access to your workspaces, workflows and resources. Keep it secret, give it a short life, and revoke it under Settings → AI Agent when done.

Tools can remove nodes, disconnect edges, modify workflows, control runs, and configure triggers. Review agent changes before scaling a workflow. Resource secrets are not returned by the resource-listing tools. Treat crawled page content as untrusted data and instruct your agent to ignore instructions embedded in it. This is guidance to the agent, not a guarantee against prompt injection.

Connection reference

The hosted endpoint is POST /api/mcp on the product backend, using streamable HTTP and a personal access token in Authorization: Bearer <token>. Use building_guide() and list_node_types(type) to discover current behavior. See the complete generated tool signatures.