API-Level MCP Server Recipe

Purpose

Add a Model Context Protocol (MCP) endpoint to a cloud-hosted API so AI agents can access user data without the desktop app running. This recipe handles multi-tenant isolation (multiple users with separate data), agent key authentication with bcrypt, database-backed sessions for horizontal scaling, and permission enforcement on every tool call.

The core value is in the integration glue: how the MCP TypeScript SDK connects to a Bun/Elysia server, how agent keys are validated efficiently with prefix prefiltering, how sessions cache resolved permissions, and how the two-layer auth model (OAuth for user identity + agent keys for permissions) works together.

When to Use

•You have a cloud-hosted API with user data you want AI agents to access
•Multiple users need isolated MCP access to their own data
•You need agent key authentication with role-based permissions
•You want agents to work without the desktop app running
•Your API already has an auth system (Better Auth, NextAuth, Lucia, etc.)

Technology Stack

Layer	Technology	Version
Runtime	Bun	1.2+
Framework	Elysia	1.4+
MCP SDK	@modelcontextprotocol/sdk	1.26+
ORM	Drizzle ORM with postgres.js	0.45+
Database	PostgreSQL	15+
Key Hashing	bcryptjs	2.4+
Validation	Zod 4	4.x

Prerequisite Recipes:

•Elysia + BetterAuth API — Base API server setup
•BetterAuth OAuth Provider — OAuth 2.1 for transport-level authentication (required for production multi-tenant deployments)

Architecture Overview

code

MCP Client (Claude Code, Cursor, etc.)
    │
    │  HTTP + Bearer token (OAuth access_token)
    ▼
┌─────────────────────────────────────────────────┐
│  Elysia Route Handler (/api/mcp)                │
│  ├─ POST — Client-to-server MCP messages        │
│  ├─ GET  — Server-to-client SSE stream          │
│  └─ DELETE — Session termination                │
│                                                 │
│  1. OAuth middleware resolves user identity      │
│  2. 401 + WWW-Authenticate if unauthenticated   │
│  3. Create per-request MCP server + transport    │
│  4. Delegate to MCP SDK                         │
└──────────────────┬──────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────┐
│  MCP Server (stateless, per-request)            │
│                                                 │
│  Tool: authenticate                             │
│  ├─ Validates agent key (bcrypt + prefix)       │
│  ├─ Cross-validates: key owner == OAuth user    │
│  └─ Creates database session with permissions   │
│                                                 │
│  Tool: your_tool_name                           │
│  ├─ Validates session (not expired)             │
│  ├─ Enforces permission (role check)            │
│  ├─ Enforces resource access (folder check)     │
│  └─ Calls service layer with ownerId filter     │
└──────────────────┬──────────────────────────────┘
                   │
                   ▼
┌─────────────────────────────────────────────────┐
│  Service Layer                                  │
│  Every method takes ownerId as parameter        │
│  Every query includes ownerId WHERE clause      │
└──────────────────┬──────────────────────────────┘
                   │
                   ▼
              PostgreSQL

Key Design Decisions

Two-layer authentication. OAuth establishes user identity at the transport level (which human user is making the request). Agent keys establish permissions at the application level (which agent, with what role, accessing which folders). This separation means the MCP endpoint uses the same OAuth infrastructure as the rest of your API, while agent-specific permissions live in their own system.

Stateless server instances. A new MCP server is created for every HTTP request. No in-memory state, no connection lifecycle management, no cleanup. Sessions live in PostgreSQL. This is simpler and scales horizontally.

Database-backed sessions. Sessions cache resolved permissions (which tools the agent can use, which folders it can access) so each tool call doesn't need to re-resolve the full agent → membership → role → access group → folder chain. Sessions expire after 24 hours.

Prefix prefiltering for bcrypt. bcrypt with 12 rounds takes ~250ms per comparison. Without prefiltering, validating one key against 100 memberships = 25 seconds. Storing the first 8 characters of each key as a key_prefix column lets you filter candidates with a fast SQL query before running bcrypt. This reduces bcrypt calls from O(n) to O(1-2).

ownerId on everything. Every table has an owner_id column. Every service method takes ownerId as a parameter. Every query includes it in the WHERE clause. This is the multi-tenant isolation boundary — miss one and you leak data between users.

Data Model

Six tables for the MCP auth system. All tables have owner_id for multi-tenant isolation.

code

┌──────────────┐     ┌──────────────────────┐     ┌─────────────┐
│  mcp_agents  │     │  mcp_agent_group_    │     │  mcp_roles  │
│              │────▶│  membership          │◀────│             │
│  id          │     │                      │     │  id         │
│  name        │     │  id                  │     │  name       │
│  description │     │  agent_id       (FK) │     │  permissions│ ← JSON array
│  owner_id    │     │  access_group_id(FK) │     │  owner_id   │
└──────────────┘     │  role_id        (FK) │     └─────────────┘
                     │  key_hash            │
                     │  key_prefix          │ ← first 8 chars
                     │  revoked_at          │ ← soft revocation
                     │  owner_id            │
                     └──────────┬───────────┘
                                │
                     ┌──────────▼───────────┐     ┌────────────────────────┐
                     │  mcp_access_groups   │     │  mcp_group_folder_     │
                     │                      │────▶│  access                │
                     │  id                  │     │                        │
                     │  name                │     │  access_group_id  (FK) │
                     │  description         │     │  group_id         (FK) │ ← your folder table
                     │  owner_id            │     │  owner_id              │
                     └──────────────────────┘     └────────────────────────┘

                     ┌──────────────────────┐
                     │  mcp_sessions        │
                     │                      │ ← caches resolved permissions
                     │  id                  │
                     │  agent_id            │
                     │  membership_id       │
                     │  access_group_id     │
                     │  allowed_group_ids   │ ← jsonb array of folder IDs
                     │  permissions         │ ← jsonb array of tool names
                     │  owner_id            │
                     │  expires_at          │ ← 24 hours
                     └──────────────────────┘

Key Schema Details

typescript

import {
  pgTable,
  text,
  timestamp,
  index,
  uniqueIndex,
  jsonb,
} from "drizzle-orm/pg-core";

// Agent identities
export const mcpAgents = pgTable(
  "mcp_agents",
  {
    id: text("id").primaryKey(),
    name: text("name").notNull(),
    description: text("description"),
    ownerId: text("owner_id").notNull(),
    createdAt: timestamp("created_at", { withTimezone: true }).notNull(),
    updatedAt: timestamp("updated_at", { withTimezone: true }).notNull(),
  },
  (table) => [index("idx_mcp_agents_owner").on(table.ownerId)]
);

// Agent credentials — links agents to access groups with roles
export const mcpAgentGroupMembership = pgTable(
  "mcp_agent_group_membership",
  {
    id: text("id").primaryKey(),
    agentId: text("agent_id").notNull(),
    accessGroupId: text("access_group_id").notNull(),
    roleId: text("role_id").notNull(),
    keyHash: text("key_hash").notNull(), // bcrypt hash
    keyPrefix: text("key_prefix").notNull(), // first 8 chars for prefiltering
    key: text("key").notNull().default(""), // full key (shown to user once)
    ownerId: text("owner_id").notNull(),
    createdAt: timestamp("created_at", { withTimezone: true }).notNull(),
    revokedAt: timestamp("revoked_at", { withTimezone: true }), // soft revoke
    lastUsedAt: timestamp("last_used_at", { withTimezone: true }),
  },
  (table) => [
    uniqueIndex("idx_mcp_membership_key_hash").on(table.keyHash),
    index("idx_mcp_membership_owner").on(table.ownerId),
    uniqueIndex("idx_mcp_membership_unique").on(
      table.agentId,
      table.accessGroupId
    ),
  ]
);

// Roles — named permission sets
export const mcpRoles = pgTable("mcp_roles", {
  id: text("id").primaryKey(),
  name: text("name").notNull(),
  description: text("description"),
  permissions: text("permissions").notNull(), // JSON array: ["browse", "read_document", ...]
  ownerId: text("owner_id").notNull(),
  createdAt: timestamp("created_at", { withTimezone: true }).notNull(),
  updatedAt: timestamp("updated_at", { withTimezone: true }).notNull(),
});

// Sessions — database-backed, caches resolved permissions
export const mcpSessions = pgTable(
  "mcp_sessions",
  {
    id: text("id").primaryKey(),
    agentId: text("agent_id").notNull(),
    membershipId: text("membership_id").notNull(),
    accessGroupId: text("access_group_id").notNull(),
    allowedGroupIds: jsonb("allowed_group_ids").notNull().$type<string[]>(),
    permissions: jsonb("permissions").notNull().$type<string[]>(),
    ownerId: text("owner_id").notNull(),
    createdAt: timestamp("created_at", { withTimezone: true }).notNull(),
    expiresAt: timestamp("expires_at", { withTimezone: true }).notNull(),
  },
  (table) => [index("idx_mcp_sessions_expires").on(table.expiresAt)]
);

Key constraints:

•key_hash has a UNIQUE index — prevents key reuse across memberships
•(agent_id, access_group_id) has a UNIQUE index — one membership per agent-group pair
•permissions uses jsonb — no schema migration needed when adding tools
•revoked_at enables soft revocation with an audit trail

Service Layer

Permission Constants

Define a central registry of permission names that map 1:1 with tool names. This registry is used for role definitions, session caching, and enforcement.

typescript

// permissions/constants.ts
export const MCP_PERMISSIONS = {
  // Define your tool permissions here
  // Example read permissions:
  browse: "browse",
  search: "search",
  read_item: "read_item",
  // Example write permissions:
  create_item: "create_item",
  update_item: "update_item",
  delete_item: "delete_item",
} as const;

export type McpPermission =
  (typeof MCP_PERMISSIONS)[keyof typeof MCP_PERMISSIONS];

export const ALL_PERMISSIONS = Object.values(MCP_PERMISSIONS);

Permission Enforcement

typescript

// permissions.ts
export function enforcePermission(
  session: { permissions: string[] },
  permission: string,
  toolName: string
): void {
  const perms = session.permissions as string[];
  if (!perms.includes(permission)) {
    throw new Error(
      `Permission denied: tool "${toolName}" requires "${permission}" permission. ` +
        `Your permissions are: ${perms.join(", ")}`
    );
  }
}

Subtree Folder Access Check

If an agent has access to folder "Projects", it implicitly has access to "Projects/Work/Q1". The check walks up the parent chain from the target folder to see if any ancestor is in the allowed list.

typescript

// permissions.ts
export async function isGroupOrAncestorAllowed(
  groupId: string,
  allowedGroupIds: string[],
  ownerId: string
): Promise<boolean> {
  const allowedSet = new Set(allowedGroupIds);
  const visited = new Set<string>(); // Circular reference guard
  let currentGroupId: string | null = groupId;

  while (currentGroupId) {
    if (visited.has(currentGroupId)) return false;
    visited.add(currentGroupId);

    if (allowedSet.has(currentGroupId)) return true;

    // Walk up to parent — your service provides this
    currentGroupId = await getParentGroupId(currentGroupId, ownerId);
  }

  return false;
}

Agent Key Validation (Critical Path)

The 3-stage validation flow with prefix prefiltering:

typescript

// services/membership.service.ts
import bcrypt from "bcryptjs";

export const membershipService = {
  async validateAgentKey(agentKey: string) {
    // Stage 1: Basic validation
    if (!agentKey || typeof agentKey !== "string" || agentKey.length < 10) {
      return null;
    }

    // Stage 2: Prefix prefiltering (CRITICAL for performance)
    const prefix = agentKey.substring(0, 8);
    const candidates = await db
      .select()
      .from(mcpAgentGroupMembership)
      .where(
        and(
          eq(mcpAgentGroupMembership.keyPrefix, prefix),
          isNull(mcpAgentGroupMembership.revokedAt) // Exclude revoked
        )
      );

    // Stage 3: bcrypt compare — slow but secure, only 1-2 candidates
    let membership = null;
    for (const candidate of candidates) {
      if (await bcrypt.compare(agentKey, candidate.keyHash)) {
        membership = candidate;
        break;
      }
    }
    if (!membership) return null;

    // Load related entities (agent, access group, role, folder access)
    const [agent] = await db
      .select()
      .from(mcpAgents)
      .where(eq(mcpAgents.id, membership.agentId))
      .limit(1);
    if (!agent) return null;

    const [accessGroup] = await db
      .select()
      .from(mcpAccessGroups)
      .where(eq(mcpAccessGroups.id, membership.accessGroupId))
      .limit(1);
    if (!accessGroup) return null;

    const [role] = await db
      .select()
      .from(mcpRoles)
      .where(eq(mcpRoles.id, membership.roleId))
      .limit(1);
    if (!role) return null;

    const permissions = JSON.parse(role.permissions) as string[];

    const folderAccess = await db
      .select({ groupId: mcpGroupFolderAccess.groupId })
      .from(mcpGroupFolderAccess)
      .where(eq(mcpGroupFolderAccess.accessGroupId, accessGroup.id));
    const allowedFolderIds = folderAccess.map((fa) => fa.groupId);

    return {
      membership,
      agent,
      accessGroup,
      role,
      allowedFolderIds,
      permissions,
      ownerId: membership.ownerId,
    };
  },

  async updateLastUsed(membershipId: string): Promise<void> {
    await db
      .update(mcpAgentGroupMembership)
      .set({ lastUsedAt: new Date() })
      .where(eq(mcpAgentGroupMembership.id, membershipId));
  },
};

Session Service

typescript

// services/session.service.ts
const SESSION_TTL_HOURS = 24;
const HOUR_IN_MS = 60 * 60 * 1000;

export const sessionService = {
  async createSession(
    agentId: string,
    membershipId: string,
    accessGroupId: string,
    allowedGroupIds: string[],
    permissions: string[],
    ownerId: string
  ): Promise<{ sessionId: string; expiresAt: Date }> {
    const sessionId = crypto.randomUUID(); // or nanoid()
    const now = new Date();
    const expiresAt = new Date(now.getTime() + SESSION_TTL_HOURS * HOUR_IN_MS);

    await db.insert(mcpSessions).values({
      id: sessionId,
      agentId,
      membershipId,
      accessGroupId,
      allowedGroupIds,
      permissions,
      ownerId,
      createdAt: now,
      expiresAt,
    });

    return { sessionId, expiresAt };
  },

  async getSession(sessionId: string) {
    const [session] = await db
      .select()
      .from(mcpSessions)
      .where(eq(mcpSessions.id, sessionId))
      .limit(1);

    if (!session) return null;
    if (new Date() > session.expiresAt) {
      await db.delete(mcpSessions).where(eq(mcpSessions.id, sessionId));
      return null;
    }

    return session;
  },

  async deleteSession(sessionId: string): Promise<void> {
    await db.delete(mcpSessions).where(eq(mcpSessions.id, sessionId));
  },

  async cleanupExpiredSessions(): Promise<number> {
    const result = await db
      .delete(mcpSessions)
      .where(lt(mcpSessions.expiresAt, new Date()))
      .returning({ id: mcpSessions.id });
    return result.length;
  },
};

Implementation Process

Phase 1: Install Dependencies and Schema

1.1 Install the MCP SDK and bcrypt

bash

bun add @modelcontextprotocol/sdk bcryptjs zod
bun add -d @types/bcryptjs

IMPORTANT — Zod version: MCP SDK 1.26+ requires Zod 4. If your project is on Zod 3, upgrade to Zod 4 first. They are different packages and Zod 4 is a monorepo-wide change — budget time for it.

1.2 Create directory structure

bash

mkdir -p src/features/mcp/services
mkdir -p src/routes/mcp

1.3 Create the database schema

Create src/features/mcp/db.ts with the six tables from the Data Model section. Adapt the table and column names to your project conventions, but preserve:

•owner_id on every table
•key_hash + key_prefix on the membership table
•jsonb for allowed_group_ids and permissions on sessions
•revoked_at for soft revocation on memberships

Add the schema file to your drizzle.config.ts and run the migration:

bash

bun run db:generate
bun run db:migrate

Validate:

• Migration runs successfully
• All six tables exist with owner_id columns
• key_hash has a unique constraint

Phase 2: Transport Adapter (If Needed)

Skip this phase if: Your MCP SDK version includes WebStandardStreamableHTTPServerTransport (SDK 1.30+). This transport works natively with Fetch API Request/Response and requires no adapter.

You need an adapter if: Your SDK version only has StreamableHTTPServerTransport (which expects Node.js IncomingMessage / ServerResponse).

The Problem: The MCP SDK historically expected Node.js HTTP primitives. Bun, Deno, Cloudflare Workers, and similar runtimes use Web Standard Request/Response. Passing a Request where the SDK expects IncomingMessage crashes on .on(), .pipe(), .writeHead(), etc.

The Adapter Pattern:

Create src/features/mcp/adapter.ts:

typescript

import { IncomingMessage, ServerResponse } from "node:http";
import { Socket } from "node:net";

/**
 * Convert Web Standard Request to Node.js IncomingMessage + ServerResponse
 * for use with StreamableHTTPServerTransport.
 */
export function toReqRes(request: Request): {
  req: IncomingMessage;
  res: MockServerResponse;
} {
  const socket = new Socket();
  const req = new IncomingMessage(socket);
  req.method = request.method;
  req.url = new URL(request.url).pathname;

  const headers: Record<string, string> = {};
  request.headers.forEach((value, key) => {
    headers[key.toLowerCase()] = value;
  });
  req.headers = headers;

  const res = new MockServerResponse(req);
  return { req, res };
}

/**
 * Convert captured MockServerResponse back to Web Standard Response.
 */
export function toFetchResponse(res: MockServerResponse): Response {
  const headers = new Headers();
  const rawHeaders = res.getHeaders();
  for (const [key, value] of Object.entries(rawHeaders)) {
    if (value !== undefined) {
      if (Array.isArray(value)) {
        for (const v of value) headers.append(key, v);
      } else {
        headers.set(key, String(value));
      }
    }
  }

  const body = res.getBody();
  const contentType = headers.get("content-type") || "";

  // Handle SSE streams
  if (contentType.includes("text/event-stream")) {
    const stream = new ReadableStream({
      start(controller) {
        if (body.length > 0) {
          controller.enqueue(new TextEncoder().encode(body));
        }
        controller.close();
      },
    });
    return new Response(stream, { status: res.statusCode, headers });
  }

  return new Response(body || null, { status: res.statusCode, headers });
}

/**
 * ServerResponse subclass that captures written data.
 */
export class MockServerResponse extends ServerResponse {
  private _chunks: Buffer[] = [];
  private _ended = false;

  constructor(req: IncomingMessage) {
    super(req);
    this.assignSocket(
      new CaptureSocket(this._chunks, () => {
        this._ended = true;
      })
    );
  }

  getBody(): string {
    return Buffer.concat(this._chunks).toString("utf-8");
  }

  hasEnded(): boolean {
    return this._ended;
  }
}

/**
 * Socket subclass that captures writes instead of sending over TCP.
 */
class CaptureSocket extends Socket {
  constructor(
    private chunks: Buffer[],
    private onEnd: () => void
  ) {
    super();
    Object.defineProperty(this, "writable", { value: true, writable: true });
  }

  override write(
    data: Uint8Array | string,
    encodingOrCb?: BufferEncoding | ((err?: Error) => void),
    cb?: (err?: Error) => void
  ): boolean {
    const encoding = typeof encodingOrCb === "string" ? encodingOrCb : "utf-8";
    const callback = typeof encodingOrCb === "function" ? encodingOrCb : cb;
    if (typeof data === "string") {
      this.chunks.push(Buffer.from(data, encoding));
    } else {
      this.chunks.push(Buffer.from(data));
    }
    if (callback) callback();
    return true;
  }

  override end(
    data?: unknown,
    encodingOrCb?: BufferEncoding | (() => void),
    cb?: () => void
  ): this {
    if (data) {
      if (typeof data === "string") {
        const encoding =
          typeof encodingOrCb === "string" ? encodingOrCb : "utf-8";
        this.chunks.push(Buffer.from(data, encoding));
      } else if (Buffer.isBuffer(data) || data instanceof Uint8Array) {
        this.chunks.push(Buffer.from(data));
      }
    }
    this.onEnd();
    const callback = typeof encodingOrCb === "function" ? encodingOrCb : cb;
    if (callback) callback();
    return this;
  }

  override destroy(): this {
    return this;
  }
}

When to use which approach:

SDK Version	Transport Class	Adapter Needed?
< 1.30	`StreamableHTTPServerTransport`	Yes
>= 1.30	`WebStandardStreamableHTTPServerTransport`	No

Validate:

• POST to /api/mcp with {"jsonrpc":"2.0","id":1,"method":"tools/list"} returns a tool list (after auth is set up)

Phase 3: MCP Server Factory and Route Handler

3.1 Server factory (src/features/mcp/server.ts)

Create a new MCP server per request. Pass the OAuth user ID for cross-validation.

typescript

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { registerTools } from "./tools";

export function createMCPServer(oauthUserId: string): McpServer {
  const server = new McpServer(
    { name: "your-app-api", version: "1.0.0" },
    { capabilities: { tools: {} } }
  );

  registerTools(server, oauthUserId);
  return server;
}

3.2 Route handler (src/routes/mcp/index.ts)

The route handler creates a per-request server + transport, delegates to the MCP SDK, and returns the response.

With WebStandardStreamableHTTPServerTransport (SDK 1.30+):

typescript

import { Elysia } from "elysia";
import { WebStandardStreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/webStandardStreamableHttp.js";
import { createMCPServer } from "@features/mcp/server";
import { sessionMiddleware } from "@core/http";
import { env } from "@core/config";
import type { User } from "@features/auth";

const RESOURCE_METADATA_URL = `${env.BETTER_AUTH_URL}/.well-known/oauth-protected-resource`;

function unauthorizedResponse(): Response {
  return new Response(
    JSON.stringify({
      jsonrpc: "2.0",
      error: { code: -32001, message: "Authentication required" },
    }),
    {
      status: 401,
      headers: {
        "content-type": "application/json",
        "www-authenticate": `Bearer resource_metadata="${RESOURCE_METADATA_URL}"`,
      },
    }
  );
}

export const mcpRoutes = new Elysia({ prefix: "/api/mcp" })
  .use(sessionMiddleware)
  .post("", async ({ request, user }) => {
    if (!user) return unauthorizedResponse();
    return handleMcpRequest(request, user);
  })
  .get("", async ({ request, user }) => {
    if (!user) return unauthorizedResponse();
    return handleMcpRequest(request, user);
  })
  .delete("", async ({ request, user }) => {
    if (!user) return unauthorizedResponse();
    return handleMcpRequest(request, user);
  });

async function handleMcpRequest(
  request: Request,
  user: User
): Promise<Response> {
  try {
    const server = createMCPServer(user.id);

    const transport = new WebStandardStreamableHTTPServerTransport({
      sessionIdGenerator: undefined, // Stateless mode
    });

    await server.connect(transport);
    const response = await transport.handleRequest(request);

    // Clean up non-streaming responses
    const contentType = response.headers.get("content-type") || "";
    if (!contentType.includes("text/event-stream")) {
      await transport.close();
      await server.close();
    }

    return response;
  } catch (err) {
    console.error("[mcp] handleMcpRequest error:", err);
    return new Response(
      JSON.stringify({
        jsonrpc: "2.0",
        error: { code: -32603, message: "Internal server error" },
      }),
      { status: 500, headers: { "content-type": "application/json" } }
    );
  }
}

With the Node.js adapter (SDK < 1.30):

Replace WebStandardStreamableHTTPServerTransport with StreamableHTTPServerTransport and use the adapter:

typescript

import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
import { toReqRes, toFetchResponse } from "@features/mcp/adapter";

async function handleMcpRequest(request: Request, user: User) {
  const server = createMCPServer(user.id);
  const transport = new StreamableHTTPServerTransport({
    sessionIdGenerator: undefined,
  });
  await server.connect(transport);

  const { req, res } = toReqRes(request);
  await transport.handleRequest(req, res);
  return toFetchResponse(res);
}

Validate:

• POST to /api/mcp without auth returns 401 with WWW-Authenticate header
• GET to /api/mcp doesn't crash (SSE endpoint)
• DELETE to /api/mcp doesn't crash (session termination)

Phase 4: Tool Registration

Tools are where your application logic lives. The recipe provides the authenticate tool (required) and the tool registration pattern that you adapt for your domain.

4.1 Session validation helper

Every tool except authenticate needs this:

typescript

// tools.ts
async function getValidatedSession(sessionId: string, oauthUserId: string) {
  const session = await sessionService.getSession(sessionId);
  if (!session) {
    throw new Error("Session expired or invalid. Please re-authenticate.");
  }
  // Cross-validate: session owner must match OAuth user
  if (session.ownerId !== oauthUserId) {
    throw new Error("Session does not belong to the authenticated user.");
  }
  return session;
}

4.2 The authenticate tool (required)

This tool exchanges an agent key for a session ID:

typescript

import { z } from "zod";
import type { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";

export function registerTools(server: McpServer, oauthUserId: string): void {
  // Always register authenticate first
  server.tool(
    "authenticate",
    "Authenticate with an agent key to get a session ID. " +
      "Must be called before any other tool.",
    { agentKey: z.string() },
    async ({ agentKey }) => {
      const result = await membershipService.validateAgentKey(agentKey);
      if (!result) throw new Error("Invalid or revoked agent key");

      // Cross-validate: key owner must match OAuth user
      if (result.ownerId !== oauthUserId) {
        throw new Error("Agent key does not belong to the authenticated user");
      }

      const { sessionId, expiresAt } = await sessionService.createSession(
        result.agent.id,
        result.membership.id,
        result.accessGroup.id,
        result.allowedFolderIds,
        result.permissions,
        result.ownerId
      );

      await membershipService.updateLastUsed(result.membership.id);

      // Tell the agent what it can and cannot do
      const deniedTools = ALL_PERMISSIONS.filter(
        (p) => !result.permissions.includes(p)
      );

      let instructions = "Use sessionId in all subsequent tool calls. ";
      instructions += `Your permissions: ${result.permissions.join(", ")}. `;
      if (deniedTools.length > 0) {
        instructions += `NOT available: ${deniedTools.join(", ")}.`;
      }

      return {
        content: [
          {
            type: "text",
            text: JSON.stringify(
              {
                sessionId,
                expiresAt: expiresAt.toISOString(),
                accessGroup: {
                  id: result.accessGroup.id,
                  name: result.accessGroup.name,
                },
                role: result.role
                  ? { id: result.role.id, name: result.role.name }
                  : null,
                permissions: result.permissions,
                instructions,
              },
              null,
              2
            ),
          },
        ],
      };
    }
  );

  // Register your domain-specific tools below
  // registerYourReadTool(server, oauthUserId);
  // registerYourWriteTool(server, oauthUserId);
}

4.3 Read tool pattern

Every read tool follows this structure:

typescript

server.tool(
  "your_read_tool",
  "Description of what this tool reads. Requires your_read_tool permission.",
  {
    sessionId: z.string(),
    // ... your tool-specific parameters
  },
  async ({ sessionId, ...params }) => {
    // 1. Validate session + cross-check OAuth user
    const session = await getValidatedSession(sessionId, oauthUserId);

    // 2. Enforce permission
    enforcePermission(session, "your_read_tool", "your_read_tool");

    // 3. Check resource access (if applicable)
    const allowedGroupIds = session.allowedGroupIds as string[];
    const isAllowed = await isGroupOrAncestorAllowed(
      targetGroupId,
      allowedGroupIds,
      session.ownerId
    );
    if (!isAllowed) throw new Error("Resource not accessible");

    // 4. Call service layer with ownerId
    const data = await yourService.get(params.id, session.ownerId);

    // 5. Return JSON result
    return {
      content: [{ type: "text", text: JSON.stringify(data, null, 2) }],
    };
  }
);

4.4 Write tool pattern

Write tools add resource existence validation and may include domain-specific logic:

typescript

server.tool(
  "your_write_tool",
  "Description. Requires your_write_tool permission and folder access.",
  {
    sessionId: z.string(),
    targetGroupId: z.string(),
    // ... your write-specific parameters
  },
  async ({ sessionId, targetGroupId, ...params }) => {
    // 1. Validate session + cross-check OAuth user
    const session = await getValidatedSession(sessionId, oauthUserId);

    // 2. Enforce permission
    enforcePermission(session, "your_write_tool", "your_write_tool");

    // 3. Validate target resource exists
    const group = await groupService.get(targetGroupId, session.ownerId);
    if (!group) throw new Error(`Folder ${targetGroupId} not found`);

    // 4. Check folder access (subtree model)
    const allowedGroupIds = session.allowedGroupIds as string[];
    const isAllowed = await isGroupOrAncestorAllowed(
      targetGroupId,
      allowedGroupIds,
      session.ownerId
    );
    if (!isAllowed) {
      throw new Error(`Folder ${targetGroupId} is not accessible`);
    }

    // 5. Perform the write via service layer
    const result = await yourService.create(
      params,
      session.ownerId,
      `ai:agent:${session.agentId}` // Track who created it
    );

    // 6. Return result
    return {
      content: [
        {
          type: "text",
          text: JSON.stringify({ id: result.id, ...result }, null, 2),
        },
      ],
    };
  }
);

Validate:

• authenticate returns session with permissions
• Tools reject expired/invalid session IDs
• Tools reject calls without required permissions
• Subtree folder access works (parent access grants child access)
• Cross-user validation prevents User A from using User B's agent key

Phase 5: OAuth Discovery Endpoints (MCP-Specific)

MCP clients (Claude Code, Cursor) discover how to authenticate by following the OAuth Protected Resource Metadata standard (RFC 9728). This is the intersection where MCP relies on OAuth — these endpoints are what make automatic OAuth discovery work.

5.1 Protected Resource Metadata (RFC 9728)

This tells MCP clients where to find the authorization server:

typescript

// src/routes/mcp/resource-metadata.ts
import { Elysia } from "elysia";
import { env } from "@core/config";

export const resourceMetadataRoute = new Elysia().get(
  "/.well-known/oauth-protected-resource",
  () => ({
    resource: `${env.BETTER_AUTH_URL}/api/mcp`,
    authorization_servers: [env.BETTER_AUTH_URL],
    bearer_methods_supported: ["header"],
    scopes_supported: ["openid", "profile", "email", "offline_access"],
  })
);

5.2 Authorization Server Metadata (RFC 8414)

If you're using BetterAuth with the OAuth Provider plugin, it provides a helper to generate this metadata:

typescript

// src/routes/mcp/auth-server-metadata.ts
import { Elysia } from "elysia";
import { oauthProviderAuthServerMetadata } from "@better-auth/oauth-provider";
import { auth } from "@features/auth";

const handler = oauthProviderAuthServerMetadata(auth);

export const authServerMetadataRoute = new Elysia().get(
  "/.well-known/oauth-authorization-server",
  async ({ request }) => handler(request)
);

IMPORTANT: These well-known endpoints must be at the domain root (per RFC 8615), not under /api/mcp. Register them on the app root:

typescript

// src/index.ts
import { resourceMetadataRoute } from "./routes/mcp/resource-metadata";
import { authServerMetadataRoute } from "./routes/mcp/auth-server-metadata";
import { mcpRoutes } from "./routes/mcp";

const app = new Elysia()
  // ... cors, session middleware, auth handler ...
  .use(resourceMetadataRoute) // /.well-known/oauth-protected-resource
  .use(authServerMetadataRoute) // /.well-known/oauth-authorization-server
  .use(mcpRoutes) // /api/mcp
  .listen({ hostname: "0.0.0.0", port: env.BUN_PORT });

How the discovery flow works:

•MCP client POSTs to /api/mcp — gets 401 with WWW-Authenticate: Bearer resource_metadata="..."
•Client fetches /.well-known/oauth-protected-resource — learns the authorization server URL
•Client fetches /.well-known/oauth-authorization-server — learns the authorize, token, and registration endpoints
•Client performs OAuth PKCE flow in a browser
•Client retries the POST with Authorization: Bearer {access_token}
•Server resolves the user from the bearer token and handles the MCP request

If your auth system isn't BetterAuth, you need to serve equivalent metadata documents from your authorization server. The specific endpoints (authorization_endpoint, token_endpoint, etc.) depend on your OAuth provider.

Validate:

• GET /.well-known/oauth-protected-resource returns valid JSON with authorization_servers
• GET /.well-known/oauth-authorization-server returns valid OAuth metadata
• Unauthenticated POST to /api/mcp returns 401 with WWW-Authenticate header containing the resource metadata URL

Phase 6: Testing

6.1 Unit tests with InMemoryTransport

The MCP SDK provides InMemoryTransport for testing without HTTP:

typescript

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { InMemoryTransport } from "@modelcontextprotocol/sdk/inMemory.js";

async function createTestClient(oauthUserId: string) {
  const server = createMCPServer(oauthUserId);
  const [clientTransport, serverTransport] =
    InMemoryTransport.createLinkedPair();
  const client = new Client({ name: "test", version: "1.0.0" });

  await Promise.all([
    client.connect(clientTransport),
    server.connect(serverTransport),
  ]);

  return client;
}

// Test categories to cover:
// 1. Authentication — valid key, invalid key, revoked key, cross-user rejection
// 2. Session management — creation, expiry, invalid session
// 3. Permission enforcement — each tool checks its required permission
// 4. Resource access — subtree model, parent chain traversal
// 5. Multi-tenant isolation — User A cannot see User B's data
// 6. Tool functionality — each tool returns correct data shape

6.2 Manual testing with curl

bash

# List tools (requires auth)
curl -X POST https://your-api.com/api/mcp \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_OAUTH_TOKEN" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

# Authenticate with agent key
curl -X POST https://your-api.com/api/mcp \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_OAUTH_TOKEN" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"authenticate","arguments":{"agentKey":"mcp_..."}}}'

6.3 Testing with Claude Code

Create .mcp.json in your project:

json

{
  "your-app": {
    "url": "https://your-api.com/api/mcp"
  }
}

Claude Code will discover the OAuth endpoints automatically from the 401 response and perform the PKCE flow in a browser.

Validate:

• Unit tests pass for all tool categories
• curl requests return expected responses
• Claude Code can connect, authenticate, and use tools
• Cross-user isolation verified with two separate accounts

Integration Points

With Your Existing Auth System

The MCP endpoint uses your existing session middleware to resolve the OAuth user. The specific integration depends on your auth library:

•BetterAuth: auth.api.getSession({ headers: request.headers }) for cookies, or route through auth.handler for OAuth bearer tokens
•NextAuth/Lucia/etc.: Use your existing session middleware — the MCP routes just need the user ID

With Your Data Model

Your service layer methods should follow this pattern:

typescript

// Every method takes ownerId — no exceptions
async getById(id: string, ownerId: string) {
  return db.select().from(items)
    .where(and(
      eq(items.id, id),
      eq(items.ownerId, ownerId),
      isNull(items.deletedAt),  // If using soft deletes
    ))
    .limit(1);
}

With Agent Key Provisioning

You need a way for users to create agents, define roles, and generate API keys. This can be:

•A page in your existing web admin/settings
•A desktop or mobile app UI
•API endpoints for programmatic management

The recipe covers the auth validation side. Key generation is:

typescript

import { randomBytes } from "crypto";
import bcrypt from "bcryptjs";

const KEY_PREFIX = "mcp_";
const KEY_LENGTH = 60; // hex characters

function generateAgentKey(): string {
  return KEY_PREFIX + randomBytes(KEY_LENGTH / 2).toString("hex");
}

// When creating a membership:
const key = generateAgentKey();
const keyHash = await bcrypt.hash(key, 12);
const keyPrefix = key.substring(0, 8);
// Store keyHash, keyPrefix in the membership row
// Show the full key to the user once — it cannot be recovered

Gotchas & Important Notes

•
MCP SDK expects Node.js HTTP on older versions. If you see TypeError: req.on is not a function or res.writeHead is not a function, you need the transport adapter from Phase 2. SDK 1.30+ includes WebStandardStreamableHTTPServerTransport which eliminates this issue.
•
bcrypt is slow without prefix prefiltering. bcrypt with 12 rounds takes ~250ms per comparison. Validating one key against 100 memberships without prefiltering = 25 seconds. Always store and filter by key_prefix.
•
Zod 4 is required for MCP SDK 1.26+. The SDK uses Zod 4 internally for tool parameter validation. If your project is on Zod 3, this is a monorepo-wide upgrade that must happen before MCP integration.
•
Sessions must be database-backed. In-memory session storage works on one server instance but fails after restart or on a different instance. Always use PostgreSQL for session storage in API servers.
•
Missing ownerId filter leaks data between users. The most dangerous bug in a multi-tenant system. Establish the convention that every service method takes ownerId and every query includes it. Write tests specifically for isolation.
•
MCP client connections can be fragile. The connection flow involves multiple round-trips (OAuth discovery, PKCE flow, MCP initialization). Proxy layers add failure modes. Support direct HTTP connections where possible.
•
Permission list may outpace tool implementation. If your auth system defines permissions like create_document but the tool doesn't exist yet, include those permissions in the denied list in the authenticate response so agents know what's not available.
•
The 401 response format matters. MCP clients parse the WWW-Authenticate header to find the resource metadata URL. The exact format is: Bearer resource_metadata="https://..." with the URL in double quotes. If this is malformed, clients can't discover your OAuth server.
•
Well-known endpoints must be at the domain root. RFC 8615 requires /.well-known/* paths at the domain root, not under /api. Register them directly on the Elysia app, not in a prefixed route group.

External Documentation

•MCP Specification: https://modelcontextprotocol.io/
•MCP TypeScript SDK: https://github.com/modelcontextprotocol/typescript-sdk
•RFC 9728 (Protected Resource Metadata): https://www.rfc-editor.org/rfc/rfc9728
•RFC 8414 (OAuth Server Metadata): https://www.rfc-editor.org/rfc/rfc8414
•bcryptjs: https://github.com/dcodeIO/bcrypt.js
•Drizzle ORM: https://orm.drizzle.team
•Elysia: https://elysiajs.com