Vantage MCP Server

About

The Vantage MCP Server exposes tools for listing, querying, and creating Vantage resources. Tools are defined in /src/tools; see the Vantage MCP documentation for the full tool list, prompting examples, and product details.

This repository supports two deployment modes:

Start with the hosted MCP unless you have a specific reason to self-host.

Related MCP server: Cost Management MCP

Setup Instructions

General

The hosted MCP server uses Streamable HTTP with OAuth 2.1. After connecting a client, a browser window prompts you to sign in to Vantage — the same account you use at console.vantage.sh.

Server URL: https://mcp.vantage.sh/mcp

Clients that support remote MCP natively can connect directly to that URL. For clients that only support stdio, use the mcp-remote bridge (see Visual Studio Code below).

API token auth: If your client or script cannot complete OAuth, pass a Vantage API token as a Bearer token instead:

Authorization: Bearer <your_vantage_api_token>

Claude Code

claude mcp add --transport http vantage https://mcp.vantage.sh/mcp

Then run /mcp in a Claude Code session to complete the OAuth flow.

Codex

codex mcp add vantage --url https://mcp.vantage.sh/mcp

Run codex mcp login vantage if Codex prompts you to authenticate.

Cursor

The recommended install is the official Vantage plugin — in the Cursor chat panel, run:

/add-plugin vantage

To configure manually (for example, to commit .cursor/mcp.json to a repo):

{
  "mcpServers": {
    "VantageMCP": {
      "url": "https://mcp.vantage.sh/mcp"
    }
  }
}

Open Cursor Settings → Tools & MCP → New MCP Server to edit mcp.json, or create .cursor/mcp.json in your project root.

Claude

Claude.ai (Team / Enterprise): Settings → Connectors → Add custom connector — enter https://mcp.vantage.sh/mcp.

Claude Desktop: Settings → Connectors — add a custom connector with the same URL.

Goose

1. Extensions → Add custom extension

2. Type: Streamable HTTP

3. Endpoint: https://mcp.vantage.sh/mcp

Complete the OAuth flow when Goose connects.

Visual Studio Code

{
  "mcpServers": {
    "vantage": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.vantage.sh/mcp"]
    }
  }
}

Or use the command palette: MCP: Add Server → Command (stdio) and enter:

npx -y mcp-remote https://mcp.vantage.sh/mcp

Windsurf

Under Cascade → MCP servers → Add custom server, add:

{
  "mcpServers": {
    "vantage": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.vantage.sh/mcp"]
    }
  }
}

Zed

In Zed settings, add:

{
  "context_servers": {
    "vantage": {
      "source": "custom",
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.vantage.sh/mcp"],
      "env": {}
    }
  }
}

Other clients

For any stdio-only MCP client, use:

• Command: npx

• Arguments: -y mcp-remote https://mcp.vantage.sh/mcp

See the MCP clients list and the Vantage MCP documentation for additional clients (including ChatGPT via the Vantage app).

Authorize

After configuring your client, you may need to restart it. A browser window opens for Vantage OAuth — sign in and click Allow Access. You can revoke access anytime under Vantage Settings → API Access Tokens → MCP Server Token.

Try a prompt like: "In Vantage, which workspaces do I have access to?"

Self-Hosted (Local) MCP

Use self-hosted mode when you need API-key auth without OAuth, or when developing against this repository.

npm package (recommended for self-host)

Install nothing — configure your MCP client to run the published package:

{
  "mcpServers": {
    "Vantage": {
      "command": "npx",
      "args": ["-y", "vantage-mcp-server"],
      "env": {
        "VANTAGE_TOKEN": "<your_vantage_api_token>"
      }
    }
  }
}

Create a token in the Vantage console. Requires Node.js 20+.

From this repository (contributors)

For working on the MCP server itself:

git clone https://github.com/vantage-sh/vantage-mcp-server
cd vantage-mcp-server
npm install

Configure your client to run the local entrypoint (replace <path_to_repository> and <personal_vantage_api_token>):

{
  "mcpServers": {
    "Vantage": {
      "command": "npx",
      "args": ["tsx", "<path_to_repository>/src/local.ts"],
      "env": { "VANTAGE_TOKEN": "<personal_vantage_api_token>" }
    }
  }
}

Or use the repo's installed binary directly:

{
  "mcpServers": {
    "Vantage": {
      "command": "<path_to_repository>/node_modules/.bin/tsx",
      "args": ["<path_to_repository>/src/local.ts"],
      "env": { "VANTAGE_TOKEN": "<personal_vantage_api_token>" }
    }
  }
}

Test from the terminal: npm run local (requires VANTAGE_TOKEN in the environment).

Local Development

To develop the hosted worker locally:

cp .dev.vars.example .dev.vars.development
# Edit .dev.vars.development — set VANTAGE_MCP_TOKEN to your API token
npm run dev

Wrangler serves the worker at http://localhost:8787 (default). Setting VANTAGE_MCP_TOKEN bypasses OAuth for local testing.

Running Evals

Each tool is tested with one direct prompt and one inferred prompt. Both runs load the target tool plus four distractors, then the same cases can be replayed across models. Outcomes are committed as JSON so we do not re-run the whole suite; GitHub Pages rebuilds the report from those files. The approach and full workflow are in evals.md.

New tools must ship with cases under evals/cases/<resource>/<tool>.eval.ts. Authoring conventions: .agents/skills/writing-evals/SKILL.md.

View outcomes locally

npm run eval:site
open evals/site/index.html

No model API keys required. The published report is at https://vantage-sh.github.io/vantage-mcp-server/.

Run a new eval

Set ANTHROPIC_API_KEY and/or OPENAI_API_KEY in .env, then:

npm run eval -- --tool <your-tool> --model gpt-5.6-sol-high

--model is required (npm run eval -- --list-models prints the catalog). That writes evals/results/<model>/<resource>/<tool>.json and leaves every other result file untouched. Commit the new JSON.

Overwrite an eval

Re-run the same --tool and --model. The JSON for that pair is replaced. Do this after changing a tool's description, schema, or prompts — you do not need to re-run every model unless you want those baselines refreshed too.

Available Scripts

• npm run dev — Start the Wrangler development server

• npm run local — Run the stdio MCP server locally

• npm run inspect — Launch the MCP inspector

• npm run test — Run Vitest

• npm run format / npm run lint:fix — Biome format and lint

• npm run type-check — TypeScript check

• npm run cf-typegen — Generate Cloudflare Worker types

• npm run generate-tools-index — Regenerate src/tools/index.ts after tool changes

• npm run generate-resources-index — Regenerate src/resources/index.ts after resource changes

• npm run eval — Run tool-selection evals (npm run eval -- --tool <name> --model gpt-5.6-sol-high)

• npm run eval:site — Merge stored JSON and write evals/site/index.html

• npm run eval:view — Open promptfoo's local results viewer

Contribution Guidelines

1. Fork this repository and create a branch: git checkout -b feature/my-feature.

2. Make your changes.

3. Verify locally:

   npm run format
   npm run lint:fix
   npm run type-check
   npm test -- --run

4. Submit a pull request.

See AGENTS.md for conventions when adding tools, evals, or resources.

License

See LICENSE.md for commercial and non-commercial licensing details.