@vibelogin/mcp

@vibelogin/mcp

VibeLogin MCP server — zero-touch auth setup for AI coding agents (Cursor, Claude Desktop, Claude Code, Windsurf, Cline, Zed, …).

Spin up a VibeLogin project, configure auth methods, wire up Google OAuth, and scaffold a working sign-in flow into your codebase — all from inside your editor's chat, without leaving the IDE.

What it does

This is a Model Context Protocol server that exposes VibeLogin as a set of tools any MCP-capable agent can call. The agent decides when to use them based on the conversation; you just say "add login to my app" and it takes care of the rest.

The server runs as a local subprocess over stdio, talks to the VibeLogin console via a loopback OAuth flow (RFC 8252 + PKCE S256), and caches a refresh token under ~/.vibelogin/credentials.json so you only authorize once per month.

Install

The server runs with bunx (or npx once published) — no global install needed.

bunx @vibelogin/mcp           # bun
npx  @vibelogin/mcp           # node

You'll wire that command into your MCP client config (see below). The first tool call opens a browser for one-click consent.

Client configuration

Claude Desktop / Claude Code

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on your OS:

{
  "mcpServers": {
    "vibelogin": {
      "command": "bunx",
      "args": ["@vibelogin/mcp"],
      "env": {
        "VIBELOGIN_API_URL": "https://api.vibelogin.com",
        "VIBELOGIN_CONSOLE_URL": "https://app.vibelogin.com"
      }
    }
  }
}

Cursor

~/.cursor/mcp.json (or per-project .cursor/mcp.json):

{
  "mcpServers": {
    "vibelogin": {
      "command": "bunx",
      "args": ["@vibelogin/mcp"],
      "env": {
        "VIBELOGIN_API_URL": "https://api.vibelogin.com",
        "VIBELOGIN_CONSOLE_URL": "https://app.vibelogin.com"
      }
    }
  }
}

Windsurf / Cline / Zed

Same shape — every MCP client accepts a command + args + optional env map. Drop the snippet above into the client's MCP settings file.

Local development against your own console

Point at localhost:

"env": {
  "VIBELOGIN_API_URL": "http://localhost:3000",
  "VIBELOGIN_CONSOLE_URL": "http://localhost:3002"
}

Environment variables

Variable	Default	Purpose
VIBELOGIN_API_URL	http://localhost:3000	Hono API base URL — what the platform tools talk to.
VIBELOGIN_CONSOLE_URL	http://localhost:3002	Console base URL — used for the loopback OAuth flow.
VIBELOGIN_HOME	~/.vibelogin	Where credentials are cached.

All optional — sensible defaults for local dev. Set these in your MCP client config to point at staging or production.

Authentication flow

• The agent calls a tool that needs a token.
• If ~/.vibelogin/credentials.json has a valid (unexpired) access token, it's used as-is.
• If the access token is expired but a refresh token is present, the server calls /api/agent/refresh (no browser, ~50ms).
• If neither works, the server runs the full loopback OAuth flow:
  • Spawns a one-shot HTTP listener on 127.0.0.1:<random>
  • Generates a PKCE S256 challenge
  • Opens https://app.vibelogin.com/agent/authorize?… in your browser
  • You click "Approve" once
  • Receives the callback, exchanges the code, caches access + refresh tokens
• Subsequent tool calls reuse the cached token. Refresh tokens last 30 days; access tokens last 1 hour.

The credentials file is chmod 600 and rejected if the cached consoleUrl doesn't match the current VIBELOGIN_CONSOLE_URL (so switching between staging and prod can't reuse the wrong token).

How agents decide which tool to call

Every tool definition contains an extensive description field with explicit step number / preconditions / call-before / call-after hints — the LLM running inside Cursor / Claude Desktop / Windsurf reads those when it's deciding what to do, you don't configure anything yourself.

The standard onboarding sequence is strictly ordered

For the most common request — "add VibeLogin to my app" — the four data-modifying tools are sequential and must run in this order:

STEP 1            STEP 2              STEP 3              STEP 4
list_projects ──► create_project ──► configure_auth ──► add_auth_to_project
   (read)           (write)             (write)              (write)
   always           skip if a           enable methods        scaffold code
   first            match exists        + Google OAuth        last

Why each step depends on the previous:

Step	Why it must come first
1 → 2	Without list_projects you can't tell whether to create or reuse — running create_project blindly produces duplicates.
2 → 3	configure_auth needs a projectId that only exists after creation (or lookup).
3 → 4	add_auth_to_project scaffolds UI for whichever methods are enabled server-side. Skipping step 3 produces a sign-in page that hits METHOD_DISABLED at runtime.
4 = end	After step 4 the user has working code; no further VibeLogin tools should be called unless they ask for changes.

Re-entry points for follow-up requests (you don't restart at step 1):

User says	Enter at	Notes
"What projects do I have?"	step 1 only	Read-only, stop after.
"I already created one in the dashboard, wire it up"	get_project → step 4	Skip 1 + 2 + 3 if methods are already set.
"Turn on magic link"	step 3	Need projectId — list_projects first if unknown.
"Hook up Google"	step 3	First call without google to get the redirect URI; second call with credentials.
"Add VibeLogin to a NEW app"	step 1	Full sequence.

auth is implicit — almost never call it manually

Every other tool calls requireToken() internally, which transparently runs cached → refresh → browser-loopback with no agent involvement. Only invoke auth directly when:

• The user explicitly says "log in", "sign in", "authenticate", or "switch accounts".
• A previous tool returned code: "NOT_AUTHENTICATED" and you want to retry interactively.

Tools

The server advertises six tools via tools/list. Agents pick which to call based on your conversation.

auth

Manually run the loopback OAuth flow. You normally won't call this — requireToken triggers it automatically when needed.

Param	Type	Required	Description
clientName	string	no	Label shown on the consent screen (default "VibeLogin MCP").
noBrowser	boolean	no	Print the authorize URL to stderr instead of opening a browser. Useful for SSH / headless environments.

create_project

Creates a new VibeLogin project in your org. Returns the project's id, slug, oauthSlug, environment, publishable key, and secret key (shown once).

Param	Type	Required	Description
name	string	yes	Human-readable project name.
slug	string	no	URL-safe slug; auto-generated from name if omitted.
environment	enum	no	production | development | staging (default production).

list_projects

Lists every project in your org. Use this when the agent needs to discover what already exists before creating something new.

No parameters.

get_project

Look up a single project by id, slug, or oauthSlug.

Param	Type	Required	Description
idOrSlug	string	yes	Project id, slug, or oauthSlug.

configure_auth

Toggle auth methods, set redirect URLs, and/or wire up Google OAuth. Field-level updates — anything you don't set is left untouched.

Param	Type	Required	Description
projectId	string	yes	The project's id.
methods.emailPassword	boolean	no	Enable email + password sign-in.
methods.magicLink	boolean	no	Enable magic link emails.
methods.emailOtp	boolean	no	Enable 6-digit email OTP codes.
methods.passwordReset	boolean	no	Enable password reset emails.
methods.emailVerification	boolean	no	Require email verification on signup.
google.clientId	string	with google	Google OAuth client ID (from Google Cloud Console).
google.clientSecret	string	with google	Google OAuth client secret.
google.enabled	boolean	no	Default true.
redirectUrls	string[]	no	Overwrites the project's allowed redirect URL list.

If you call this without google, the response includes the exact Google Cloud Console URL + redirect URI you need to set up the credentials.

add_auth_to_project

Scaffold a working sign-in flow into your codebase. Detects the framework from package.json and writes non-destructive files (existing files are skipped, never clobbered).

Supported frameworks today: Next.js (App Router). Refused with guidance: Remix, Astro, Express, Vite, unknown.

Param	Type	Required	Description
projectDir	string	yes	Absolute path to the project root.
slug	string	yes	The project's oauthSlug (from create_project).
publishableKey	string	yes	The project's publishable key.
appUrl	string	yes	Your app's origin (e.g. https://myapp.com).

Files written for Next.js

middleware.ts                       # hostedAuthMiddleware
app/auth/callback/route.ts          # createCallbackHandler
app/login/page.tsx                  # <VibeLogin />
.env.local.example                  # publishable key + secret placeholder

Next steps printed by the tool:

• pnpm add @vibelogin/nextjs
• Copy .env.local.example → .env.local, fill VIBELOGIN_SECRET_KEY
• pnpm dev → open /login

Error handling

Tool failures are returned as MCP isError: true results with a structured payload so agents can branch on the failure mode instead of regex-matching free text:

{
  "isError": true,
  "structuredContent": {
    "code": "UNSUPPORTED_FRAMEWORK",
    "message": "Automated scaffolding currently supports Next.js App Router. …",
    "details": { "detected": "remix", "reason": "found `@remix-run/*` in dependencies" }
  },
  "content": [
    { "type": "text", "text": "[UNSUPPORTED_FRAMEWORK] Automated scaffolding…" }
  ]
}

Current error codes:

Code	When
NOT_AUTHENTICATED	Loopback flow ran but no valid token was persisted.
STATE_MISMATCH	OAuth callback state did not match — possible CSRF, aborted.
UNSUPPORTED_FRAMEWORK	add_auth_to_project could not detect a supported framework.
INTERNAL_ERROR	Anything else (details not populated).

API-side failures (api_error: <code> — <msg>) are still returned as INTERNAL_ERROR; they will get their own codes in a future release.

Example agent conversations

You: "Add VibeLogin to this Next.js app. Email + password and Google sign-in."

Agent: (calls create_project { name: "myapp", environment: "production" }) Created myapp (pk_live_…). (calls configure_auth { projectId, methods: { emailPassword: true } }) Email + password enabled. To turn on Google, I need a Client ID and Secret from Google Cloud Console — here's the redirect URI to authorize: https://api.vibelogin.com/oauth/myapp-xyz/google/callback. (after you paste creds:) (calls configure_auth { google: { clientId, clientSecret } }) (calls add_auth_to_project { projectDir: ".", slug: "myapp-xyz", publishableKey: "pk_live_…", appUrl: "https://myapp.com" }) Scaffold complete (next). Wrote middleware.ts, app/auth/callback/route.ts, app/login/page.tsx, .env.local.example. Run pnpm add @vibelogin/nextjs && pnpm dev and open /login.

You: "What projects do I have already?"

Agent: (calls list_projects) You have 3: myapp (production), myapp-staging (staging), internal-tools (development).

Architecture

┌──────────────┐  stdio JSON-RPC  ┌────────────────┐
│  IDE / Chat  │ ───────────────► │ @vibelogin/mcp │
│  (Cursor,    │ ◄─────────────── │   subprocess   │
│   Claude…)   │                  └────────┬───────┘
└──────────────┘                           │
                                           │ HTTPS
                                           ▼
                                  ┌─────────────────┐
                                  │ console         │
                                  │ /agent/authorize│  ◄── browser loopback
                                  │ /api/agent/token│      (PKCE S256)
                                  │ /api/agent/refresh│
                                  └────────┬────────┘
                                           │ issues HS256 platform JWT
                                           ▼
                                  ┌─────────────────┐
                                  │ Hono API        │
                                  │ /v1/platform/*  │
                                  └─────────────────┘

• Transport: stdio, one JSON-RPC 2.0 message per line.
• Logging: stderr only (stdout would corrupt the protocol).
• Zero deps: no @modelcontextprotocol/sdk — the protocol subset we use is small and stable, and zero deps means bunx @vibelogin/mcp is instant.

Security

• Loopback OAuth + PKCE S256 (RFC 8252 + RFC 7636) — no client secret embedded in the agent.
• Single-use authorization codes with 2-minute TTL, deleted on first lookup (no brute-force window for the verifier).
• Rotated refresh tokens — every successful refresh invalidates the old refresh token. Replay of a leaked token after legitimate use will fail.
• chmod 600 on ~/.vibelogin/credentials.json.
• Console-URL scoping: the cached token is rejected if the configured VIBELOGIN_CONSOLE_URL doesn't match the URL it was issued against. Switching between staging and prod cannot accidentally reuse the wrong token.
• CSRF protection: loopback callbacks must echo the OAuth state value the server generated; mismatches abort with STATE_MISMATCH.

Development

# from the monorepo root
pnpm install
pnpm --filter @vibelogin/mcp build      # tsc check (no emit)

# run the server directly (talks to local console + api)
bun packages/mcp/src/index.ts

# run the unit tests
cd packages/mcp && bun test

Tests cover the loopback flow (PKCE, URL builder, listener success/error/missing/timeout) and the stdio dispatcher.

License

Apache-2.0