Build With VULK From Claude and Other MCP Clients: The Connector, Tool by Tool

Verified on October 10, 2026, against the live server and the code behind it.

VULK runs a remote MCP server. Add one address to Claude, Claude Code or another assistant that supports remote MCP servers, sign in to VULK once, and the assistant can list your projects, read their files, check your credits, and start or continue a build for you. You do not install anything or paste an API key.

The address is:

https://app.vulk.dev/mcp

This post lists every tool the server exposes and shows the live responses that prove it is running.

Proof that it is live

Anyone can check the server without an account. Ask it for its sign-in metadata:

curl https://app.vulk.dev/.well-known/oauth-protected-resource

On October 10, 2026 it answered:

{
  "resource": "https://app.vulk.dev/mcp",
  "authorization_servers": ["https://app.vulk.dev"],
  "scopes_supported": ["projects.read", "usage.read", "projects.write"],
  "bearer_methods_supported": ["header"],
  "resource_name": "VULK"
}

Call the server itself without a token and it refuses with a 401 and a WWW-Authenticate header that points the client to that metadata. That header is how an MCP client discovers that it needs to sign the person in. The same server also answers at https://mcp.vulk.dev/mcp.

How the sign-in works

The connection uses standard OAuth, with the discovery and registration steps the MCP specification defines:

• The client registers itself automatically through a registration endpoint, so you never copy a client ID.
• The authorization code flow requires PKCE with S256, and clients authenticate with no secret, which is the right model for apps that run on your machine.
• Access is renewed with refresh tokens, but a connection lasts 90 days at most, however often it renews. After that you approve it again.
• The server speaks the MCP protocol versions 2025-06-18, 2025-03-26 and 2024-11-05.

When you connect, VULK shows a consent screen that lists exactly what the assistant will be able to do. An assistant that asks only to read gets only the read permissions, and the screen tells you the connection cannot create, change, publish or delete anything. One caution we print on that screen too: the name shown is the name the app gave itself, which VULK cannot verify, so only approve a connection you just started.

You can see every connected app under Settings › Integrations › Connected apps and remove one at any time. It loses access immediately. You can have up to 50 connected apps and 20 approval screens waiting at once.

The nine tools

Six tools only read. Three can start, continue or stop work.

Each tool declares the permission it needs, and the server refuses a call made without it. A read-only connection is not even shown the build tools.

How a build runs from an assistant

Building takes minutes, not milliseconds, so it is split in two:

1. The assistant calls create_project with your brief, up to 8,000 characters, and a request_id, a fresh UUID for this one action.
2. VULK checks your plan and credits and answers with a receipt: the request id, the project id and a link to the editor. A receipt means the build started, not that it finished.
3. The assistant calls get_operation with the same request_id until the work is done, passing on any question VULK asks you.

If a response is lost, the assistant retries with the same request_id and the same arguments, and VULK treats it as the same request instead of starting and charging a second one. If VULK has not confirmed a request within about 20 seconds, the answer says unconfirmed and the assistant reads it back with get_operation rather than guessing.

The project that comes out is an ordinary VULK project. It appears in your account, and you can open it in the editor at any time.

Built-in safeguards

Some lines are drawn on purpose, in the code and not only in the docs:

• No tool publishes a site. Publishing stays in the editor, where you can see what goes live.
• No tool buys credits or changes your plan. create_project and continue_project spend credits your plan already includes, exactly as the same request typed in the editor would. If the balance cannot cover it, VULK refuses and says why.
• No tool deletes a project. An instruction sent with continue_project can change or remove files inside a project, as the same request typed in the editor would, but no tool removes the project itself.

Answers built for models

An assistant pays for every byte a tool returns, so the answers are bounded and say when they are cut: a project list stops at 50, a file list at 500 entries, and a single file at 200,000 characters, with a truncated flag when it happens. Every result comes twice, as text for any client and as structured data for clients on the 2025-06-18 protocol, and every tool publishes the exact shape of that data as an output schema, held to it by a test. Each tool also carries a human-readable title and read-only, destructive and idempotent hints.

Failures come back as tool results the model can read, such as "no such file, list the files first", rather than protocol errors that end the conversation.

Setting it up

Claude (web and desktop): open Customize, then Connectors, add a custom connector with the address above, then choose Connect and approve in the VULK window that opens.

Claude Code:

claude mcp add --transport http vulk https://app.vulk.dev/mcp

Then run /mcp, choose vulk and sign in when the browser opens.

Other assistants: add the same address as a remote HTTP MCP server in any client that supports remote servers with sign-in. The client opens VULK in your browser for approval. The step-by-step guide is in the MCP documentation.