MCP Server

The Dragoneye MCP server lets an MCP-compatible client (such as Cursor, Claude Code, Codex CLI, Claude Desktop, VS Code Copilot Chat, or Windsurf) create and manage your Dragoneye custom vision models on your behalf, using natural-language instructions. It implements the Model Context Protocol over streamable-HTTP at https://nexus.api.dragoneye.ai/mcp and authenticates with OAuth2.

It exposes six tools that wrap the Dragoneye Model Management REST API and cover the full lifecycle of a custom model: create, list, get, update, rename, and delete. Models built through it are standard Dragoneye custom vision models supporting zero-shot object detection together with category and attribute classification.

Install

• Endpoint: https://nexus.api.dragoneye.ai/mcp
• Transport: streamable-http
• Auth: OAuth2 (browser sign-in on first use, no API key in the config)

Claude Code

Run in your terminal:

claude mcp add --transport http dragoneye https://nexus.api.dragoneye.ai/mcp

Claude Code will prompt you to sign in on first use.

Docs: Connect Claude Code to tools via MCP.

Cursor

Config file: ~/.cursor/mcp.json (or .cursor/mcp.json to scope it to one workspace)

{
  "mcpServers": {
    "dragoneye": {
      "url": "https://nexus.api.dragoneye.ai/mcp"
    }
  }
}

Cursor handles OAuth natively. Restart Cursor, and the Dragoneye tools appear in the MCP panel.

Docs: Model Context Protocol in Cursor.

Claude Desktop

Custom connectors are managed on claude.ai and sync to Claude Desktop and Claude mobile automatically.

1. In claude.ai, open Customize → Connectors.
2. Click Add custom connector.
3. Paste https://nexus.api.dragoneye.ai/mcp as the remote MCP server URL and confirm.
4. Complete sign-in in the browser when prompted.

Custom connectors require a Pro, Max, Team, or Enterprise plan.

Docs: Get started with custom connectors using remote MCP.

Codex CLI

Config file: ~/.codex/config.toml

[mcp_servers.dragoneye]
url = "https://nexus.api.dragoneye.ai/mcp"

Or let the CLI write the entry for you:

codex mcp add dragoneye --url https://nexus.api.dragoneye.ai/mcp

Codex CLI supports streamable-HTTP natively and handles the OAuth flow on first use.

Docs: Codex MCP.

VS Code (Copilot)

Config file: .vscode/mcp.json in your workspace, or run MCP: Open User Configuration from the command palette for a user-scope entry.

{
  "servers": {
    "dragoneye": {
      "type": "http",
      "url": "https://nexus.api.dragoneye.ai/mcp"
    }
  }
}

Reload VS Code; Copilot Chat will list the Dragoneye tools and prompt for sign-in on first use.

Docs: Use MCP servers in VS Code.

Windsurf

Config file: ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "dragoneye": {
      "serverUrl": "https://nexus.api.dragoneye.ai/mcp"
    }
  }
}

Windsurf handles OAuth natively. Restart Windsurf and sign in when prompted.

Docs: Model Context Protocol in Windsurf.

Other

Any spec-compliant MCP client that supports streamable-HTTP transport with OAuth2 can point directly at https://nexus.api.dragoneye.ai/mcp. Discovery metadata is published at https://nexus.api.dragoneye.ai/.well-known/mcp/server-card.json.

If your client only supports stdio, wrap the endpoint with the mcp-remote bridge:

npx -y mcp-remote https://nexus.api.dragoneye.ai/mcp

First-time sign-in

The first time your agent invokes a Dragoneye tool, your client will open a browser tab to sign you in via OAuth. Use your Dragoneye account credentials, the same ones you'd use for the dashboard. The token is cached by your client, so you only sign in once per machine. Don't have an account yet? Sign up at the Dragoneye Playground.

What the server can do

The server exposes six tools. Your agent picks the right one based on what you ask; you usually don't need to think about the names.

Tool	What it does	Example prompt
create_model	Creates a new custom vision model. Asynchronous: returns immediately in PENDING while Dragoneye builds the model.	"Build me a model that detects pallets and classifies whether they're shrink-wrapped."
list_models	Lists your existing models, newest first, with paging.	"What Dragoneye models do I have right now?"
get_model	Fetches one model by uuid, including current status and any error details. Used to poll async create / update calls.	"Is the new model done training yet?"
update_model	Changes a model's categories or attributes. Asynchronous: re-runs model generation.	"Add a forklift_brand attribute with options Toyota, Hyster, Crown, and Other."
rename_model	Renames a model in place. Synchronous, no re-training.	"Rename my test model to pallet-detector-v1."
delete_model	Permanently deletes a model. The only way to recover from a FAILED model.	"Drop the experiment-2 model."

Model creation and updates run asynchronously on Dragoneye's side, typically taking a minute or two. Your agent will poll get_model for you and report when it's done.

Tips for working with your agent

• Be explicit the first time. The agent infers the model's categories and attributes from your prompt, so listing them clearly produces better models than a vague description.
• Models in FAILED can't be updated. Ask your agent to delete and recreate; that's the supported recovery path.
• Training takes a minute. If your agent reports PENDING, just ask it to "check on it" a moment later.
• Inspect resolved state. For complex models, ask the agent to get_model and show you the resolved categories and attributes before relying on the model.

Learn more

• Server card (machine-readable): full tool schemas and discovery metadata.
• Model Management REST API Reference: the underlying API, if you'd rather call it directly.
• Create a custom vision model: what categories and attributes actually represent and how a finished model behaves.