Build Mcp Server

Build an MCP Server

You are guiding a developer through designing and building an MCP server that works seamlessly with Claude. MCP servers come in many forms — picking the wrong shape early causes painful rewrites later. Your first job is discovery, not code.

Load Claude-specific context first. The MCP spec is generic; Claude has additional auth types, review criteria, and limits. Before answering questions or scaffolding, fetch https://claude.com/docs/llms-full.txt (the full export of the Claude connector docs) so your guidance reflects Claude's actual constraints.

Do not start scaffolding until you have answers to the questions in Phase 1. If the user's opening message already answers them, acknowledge that and skip straight to the recommendation.

---

Phase 1 — Interrogate the use case

Ask these questions conversationally (batch them into one message, don't interrogate one-at-a-time). Adapt wording to what the user has already told you.

1. What does it connect to?

2. Who will use it?

• Just me / my team, on our machines → Local stdio is acceptable (easiest to prototype)
• Anyone who installs it → Remote HTTP (strongly preferred) or MCPB (if it must be local)
• Users of Claude desktop who want UI widgets → MCP app (remote or MCPB)

3. How many distinct actions does it expose?

This determines the tool-design pattern — see Phase 3.

• Under ~15 actions → one tool per action
• Dozens to hundreds of actions (e.g. wrapping a large API surface) → search + execute pattern

4. Does a tool need mid-call user input or rich display?

• Simple structured input (pick from list, enter a value, confirm) → Elicitation — spec-native, zero UI code. Host support is rolling out (Claude Code ≥2.1.76) — always pair with a capability check and fallback. See references/elicitation.md.
• Rich/visual UI (charts, custom pickers with search, live dashboards) → MCP app widgets — iframe-based, needs @modelcontextprotocol/ext-apps. See build-mcp-app skill.
• Neither → plain tool returning text/JSON.

5. What auth does the upstream service use?

• None / API key → straightforward
• OAuth 2.0 → you'll need a remote server with CIMD (preferred) or DCR support; see references/auth.md

---

Phase 2 — Recommend a deployment model

Based on the answers, recommend one path. Be opinionated. The ranked options:

⭐ Remote streamable-HTTP MCP server (default recommendation)

A hosted service speaking MCP over streamable HTTP. This is the recommended path for anything wrapping a cloud API.

Why it wins:

• Zero install friction — users add a URL, done
• One deployment serves all users; you control upgrades
• OAuth flows work properly (the server can handle redirects, DCR, token storage)
• Works across Claude desktop, Claude Code, Claude.ai, and third-party MCP hosts

Choose this unless the server must touch the user's local machine.

→ Fastest deploy: Cloudflare Workers — references/deploy-cloudflare-workers.md (zero to live URL in two commands) → Portable Node/Python: references/remote-http-scaffold.md (Express or FastMCP, runs on any host)

Elicitation (structured input, no UI build)

If a tool just needs the user to confirm, pick an option, or fill a short form, elicitation does it with zero UI code. The server sends a flat JSON schema; the host renders a native form. Spec-native, no extra packages.

Caveat: Host support is new (Claude Code shipped it in v2.1.76; Desktop unconfirmed). The SDK throws if the client doesn't advertise the capability. Always check clientCapabilities.elicitation first and have a fallback — see references/elicitation.md for the canonical pattern. This is the right spec-correct approach; host coverage will catch up.

Escalate to build-mcp-app widgets when you need: nested/complex data, scrollable/searchable lists, visual previews, live updates.

MCP app (remote HTTP + interactive UI)

Same as above, plus UI resources — interactive widgets rendered in chat. Rich pickers with search, charts, live dashboards, visual previews. Built once, renders in Claude and ChatGPT.

Choose this when elicitation's flat-form constraints don't fit — you need custom layout, large searchable lists, visual content, or live updates.

Usually remote, but can be shipped as MCPB if the UI needs to drive a local app.

→ Hand off to the `build-mcp-app` skill.

MCPB (bundled local server)

A local MCP server packaged with its runtime so users don't need Node/Python installed. The sanctioned way to ship local servers.

Choose this when the server must run on the user's machine — it reads local files, drives a desktop app, talks to localhost services, or needs OS-level access.

→ Hand off to the `build-mcpb` skill.

Local stdio (npx / uvx) — not recommended for distribution

A script launched via npx / uvx on the user's machine. Fine for personal tools and prototypes. Painful to distribute: users need the right runtime, you can't push updates, and the only distribution channel is Claude Code plugins.

Recommend this only as a stepping stone. If the user insists, scaffold it but note the MCPB upgrade path.

---

Phase 3 — Pick a tool-design pattern

Every MCP server exposes tools. How you carve them matters more than most people expect — tool schemas land directly in Claude's context window.

Pattern A: One tool per action (small surface)

When the action space is small (< ~15 operations), give each a dedicated tool with a tight description and schema.

create_issue    — Create a new issue. Params: title, body, labels[]
update_issue    — Update an existing issue. Params: id, title?, body?, state?
search_issues   — Search issues by query string. Params: query, limit?
add_comment     — Add a comment to an issue. Params: issue_id, body

Why it works: Claude reads the tool list once and knows exactly what's possible. No discovery round-trips. Each tool's schema validates inputs precisely.

Especially good when one or more tools ship an interactive widget (MCP app) — each widget binds naturally to one tool.

Pattern B: Search + execute (large surface)

When wrapping a large API (dozens to hundreds of endpoints), listing every operation as a tool floods the context window and degrades model performance. Instead, expose two tools:

search_actions  — Given a natural-language intent, return matching actions
                  with their IDs, descriptions, and parameter schemas.
execute_action  — Run an action by ID with a params object.

The server holds the full catalog internally. Claude searches, picks, executes. Context stays lean.

Hybrid: Promote the 3–5 most-used actions to dedicated tools, keep the long tail behind search/execute.

→ See references/tool-design.md for schema examples and description-writing guidance.

---

Phase 4 — Pick a framework

Recommend one of these two. Others exist but these have the best MCP-spec coverage and Claude compatibility.

If the user already has a language/stack in mind, go with it — both produce identical wire protocol.

---

Phase 5 — Scaffold and hand off

Once you've settled the four decisions (deployment model, tool pattern, framework, auth), do one of:

1. Remote HTTP, no UI → Scaffold inline using references/remote-http-scaffold.md (portable) or references/deploy-cloudflare-workers.md (fastest deploy). This skill can finish the job. 2. MCP app (UI widgets) → Summarize the decisions so far, then load the `build-mcp-app` skill. 3. MCPB (bundled local) → Summarize the decisions so far, then load the `build-mcpb` skill. 4. Local stdio prototype → Scaffold inline (simplest case), flag the MCPB upgrade path.

When handing off, restate the design brief in one paragraph so the next skill doesn't re-ask.

---

Beyond tools — the other primitives

Tools are one of three server primitives. Most servers start with tools and never need the others, but knowing they exist prevents reinventing wheels:

→ references/resources-and-prompts.md, references/elicitation.md, references/server-capabilities.md

---

Phase 6 — Test in Claude and publish

Once the server runs:

1. Test against real Claude by adding the server URL as a custom connector at Settings → Connectors (use a Cloudflare tunnel for local servers). Claude identifies itself with clientInfo.name: "claude-ai" on initialize. → https://claude.com/docs/connectors/building/testing 2. Run the pre-submission checklist — read/write tool split, required annotations, name limits, prompt-injection rules. → https://claude.com/docs/connectors/building/review-criteria 3. Submit to the Anthropic Directory. → https://claude.com/docs/connectors/building/submission 4. Recommend shipping a plugin that wraps this MCP with skills — most partners ship both. → https://claude.com/docs/connectors/building/what-to-build

---

Quick reference: decision matrix

---

Reference files

• references/remote-http-scaffold.md — minimal remote server in TS SDK and FastMCP
• references/deploy-cloudflare-workers.md — fastest deploy path (Workers-native scaffold)
• references/tool-design.md — writing tool descriptions and schemas Claude understands well
• references/auth.md — OAuth, CIMD, DCR, token storage patterns
• references/resources-and-prompts.md — the two non-tool primitives
• references/elicitation.md — spec-native user input mid-tool (capability check + fallback)
• references/server-capabilities.md — instructions, sampling, roots, logging, progress, cancellation
• references/versions.md — version-sensitive claims ledger (check when updating)

Auth for MCP Servers

Auth is the reason most people end up needing a remote server even when a local one would be simpler. OAuth redirects, token storage, and refresh all work cleanly when there's a real hosted endpoint to redirect back to.

Claude-specific authentication

Claude's MCP client supports a specific set of auth types — not every spec-compliant flow works. Full reference: https://claude.com/docs/connectors/building/authentication

Not supported: user-pasted bearer tokens (static_bearer); pure machine-to-machine client_credentials grant without user consent.

Callback URL (single, all surfaces): https://claude.ai/api/mcp/auth_callback

---

The three tiers

Tier 1: No auth / static API key

Server reads a key from env. User provides it once at setup. Done.

const apiKey = process.env.UPSTREAM_API_KEY;
if (!apiKey) throw new Error("UPSTREAM_API_KEY not set");

Works for local stdio, MCPB, and remote servers alike. If this is all you need, stop here.

Tier 2: OAuth 2.0 via CIMD (preferred per spec 2025-11-25)

Client ID Metadata Document. The MCP host publishes its client metadata at an HTTPS URL and uses that URL as its client_id. Your authorization server fetches the document, validates it, and proceeds with the auth-code flow. No registration endpoint, no stored client records.

Spec 2025-11-25 promoted CIMD to SHOULD (preferred). Advertise support via client_id_metadata_document_supported: true in your OAuth AS metadata.

Server responsibilities:

1. Serve OAuth Authorization Server Metadata (RFC 8414) at /.well-known/oauth-authorization-server with client_id_metadata_document_supported: true 2. Serve an MCP-protected-resource metadata document pointing at (1) 3. At authorize time: fetch client_id as an HTTPS URL, validate the returned client metadata, proceed 4. Validate bearer tokens on incoming /mcp requests

┌─────────┐  client_id=https://...  ┌──────────────┐   upstream OAuth   ┌──────────┐
│ MCP host│ ──────────────────────> │ Your MCP srv │ ─────────────────> │ Upstream │
└─────────┘ <─── bearer token ───── └──────────────┘ <── access token ──└──────────┘

Tier 3: OAuth 2.0 via Dynamic Client Registration (DCR)

Backward-compat fallback — spec 2025-11-25 demoted DCR to MAY. The host discovers your registration_endpoint, POSTs its metadata to register itself as a client, gets back a client_id, then runs the auth-code flow.

Implement DCR if you need to support hosts that haven't moved to CIMD yet. Same server responsibilities as CIMD, but instead of fetching the client_id URL you run a registration endpoint that stores client records.

Client priority order: pre-registered → CIMD (if AS advertises client_id_metadata_document_supported) → DCR (if AS has registration_endpoint) → prompt user.

---

Hosting providers with built-in DCR/CIMD support

Several MCP-focused hosting providers handle the OAuth plumbing for you — you implement tool logic, they run the authorization server. Check their docs for current capabilities. If the user doesn't have strong hosting preferences, this is usually the fastest path to a working OAuth-protected server.

---

Local servers and OAuth

Local stdio servers can do OAuth (open a browser, catch the redirect on a localhost port, stash the token in the OS keychain). It's fragile:

• Breaks in headless/remote environments
• Every user re-does the dance
• No central token refresh or revocation

If OAuth is required, lean hard toward remote HTTP. If you must ship local + OAuth, the @modelcontextprotocol/sdk includes a localhost-redirect helper, and MCPB is the right packaging so at least the runtime is predictable.

---

Token storage

---

Token audience validation (spec MUST)

Validating "is this a valid bearer token" isn't enough. The spec requires validating "was this token minted for this server" — RFC 8707 audience. A token issued for api.other-service.com must be rejected even if the signature checks out.

Token passthrough is explicitly forbidden. Don't accept a token, then forward it upstream. If your server needs to call another service, exchange the token or use its own credentials.

---

SDK helpers — don't hand-roll

@modelcontextprotocol/sdk/server/auth ships:

• mcpAuthRouter() — Express router for the full OAuth AS surface (metadata, authorize, token)
• bearerAuth — middleware that validates bearer tokens against your verifier
• proxyProvider — forward auth to an upstream IdP

If you're wiring auth from scratch, check these first.

Deploy to Cloudflare Workers

Fastest path from zero to a live https:// MCP URL. Free tier, no credit card to start, two commands to deploy.

Trade-off: This is a Workers-native scaffold, not a deploy target for the Express scaffold in remote-http-scaffold.md. Different runtime. If you need portability across hosts, stick with Express. If you just want it live, start here.

---

Bootstrap

npm create cloudflare@latest -- my-mcp-server \
  --template=cloudflare/ai/demos/remote-mcp-authless
cd my-mcp-server

This pulls a minimal template with the right deps (agents, zod) and a working wrangler.jsonc.

---

src/index.ts

Replace the template's calculator example with your tools. Use registerTool() (same API as the Express scaffold — the McpServer instance is identical):

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

export class MyMCP extends McpAgent {
  server = new McpServer(
    { name: "my-service", version: "0.1.0" },
    { instructions: "Prefer search_items before get_item — IDs aren't guessable." },
  );

  async init() {
    this.server.registerTool(
      "search_items",
      {
        description: "Search items by keyword. Returns up to `limit` matches.",
        inputSchema: {
          query: z.string().describe("Search keywords"),
          limit: z.number().int().min(1).max(50).default(10),
        },
        annotations: { readOnlyHint: true },
      },
      async ({ query, limit }) => {
        const results = await upstreamApi.search(query, limit);
        return { content: [{ type: "text", text: JSON.stringify(results, null, 2) }] };
      },
    );
  }
}

export default {
  fetch(request: Request, env: Env, ctx: ExecutionContext) {
    const url = new URL(request.url);
    if (url.pathname === "/mcp") {
      return MyMCP.serve("/mcp").fetch(request, env, ctx);
    }
    return new Response("Not found", { status: 404 });
  },
};

McpAgent is Cloudflare's wrapper — it handles the streamable-HTTP transport, session routing, and Durable Object plumbing. Your code only touches this.server, which is the same McpServer class from the SDK. Everything in tool-design.md and server-capabilities.md applies unchanged.

---

wrangler.jsonc

The template ships this. The Durable Objects block is boilerplate — McpAgent uses DO for session state. You don't interact with it directly.

{
  "name": "my-mcp-server",
  "main": "src/index.ts",
  "compatibility_date": "2025-03-10",
  "compatibility_flags": ["nodejs_compat"],
  "migrations": [{ "new_sqlite_classes": ["MyMCP"], "tag": "v1" }],
  "durable_objects": {
    "bindings": [{ "class_name": "MyMCP", "name": "MCP_OBJECT" }]
  }
}

If you rename the MyMCP class, update both new_sqlite_classes and class_name to match.

---

Run and deploy

npx wrangler dev     # → http://localhost:8787/mcp
npx wrangler deploy  # → https://my-mcp-server.<account>.workers.dev/mcp

wrangler deploy prints the live URL. That's the URL users paste into Claude.

Secrets (upstream API keys): npx wrangler secret put UPSTREAM_API_KEY, then read env.UPSTREAM_API_KEY inside init().

---

OAuth

Cloudflare ships @cloudflare/workers-oauth-provider — a drop-in that handles the authorization server side (CIMD/DCR endpoints, token issuance, consent UI). It wraps your McpAgent and gates /mcp behind a token check. See auth.md for the protocol details; the CF template cloudflare/ai/demos/remote-mcp-github-oauth shows the wiring.

Elicitation — spec-native user input

Elicitation lets a server pause mid-tool-call and ask the user for structured input. The client renders a native form (no iframe, no HTML). User fills it, server continues.

This is the right answer for simple input. Widgets (build-mcp-app) are for when you need rich UI — charts, searchable lists, visual previews. If you just need a confirmation, a picked option, or a few form fields, elicitation is simpler, spec-native, and works in any compliant host.

---

⚠️ Check capability first — support is new

Host support is very recent:

The SDK throws `CapabilityNotSupported` if the client doesn't advertise elicitation. There is no graceful degradation built in. You MUST check and have a fallback.

The canonical pattern

server.registerTool("delete_all", {
  description: "Delete all items after confirmation",
  inputSchema: {},
}, async ({}, extra) => {
  const caps = server.getClientCapabilities();
  if (caps?.elicitation) {
    const r = await server.elicitInput({
      mode: "form",
      message: "Delete all items? This cannot be undone.",
      requestedSchema: {
        type: "object",
        properties: { confirm: { type: "boolean", title: "Confirm deletion" } },
        required: ["confirm"],
      },
    });
    if (r.action === "accept" && r.content?.confirm) {
      await deleteAll();
      return { content: [{ type: "text", text: "Deleted." }] };
    }
    return { content: [{ type: "text", text: "Cancelled." }] };
  }
  // Fallback: return text asking Claude to relay the question
  return { content: [{ type: "text", text: "Confirmation required. Please ask the user: 'Delete all items? This cannot be undone.' Then call this tool again with their answer." }] };
});

# fastmcp
from fastmcp import Context
from fastmcp.exceptions import CapabilityNotSupported

@mcp.tool
async def delete_all(ctx: Context) -> str:
    try:
        result = await ctx.elicit("Delete all items? This cannot be undone.", response_type=bool)
        if result.action == "accept" and result.data:
            await do_delete()
            return "Deleted."
        return "Cancelled."
    except CapabilityNotSupported:
        return "Confirmation required. Ask the user to confirm deletion, then retry."

---

Schema constraints

Elicitation schemas are deliberately limited — keep forms simple:

• Flat objects only — no nesting, no arrays of objects
• Primitives only — string, number, integer, boolean, enum
• String formats limited to: email, uri, date, date-time
• Use title and description on each property — they become form labels

If your data doesn't fit these constraints, that's the signal to escalate to a widget.

---

Three-state response

Treat decline and cancel differently if it matters — decline is intentional, cancel might be accidental.

The TS SDK's server.elicitInput() auto-validates accept responses against your schema via Ajv. fastmcp's ctx.elicit() returns a typed discriminated union (AcceptedElicitation[T] | DeclinedElicitation | CancelledElicitation).

---

fastmcp response_type shorthand

await ctx.elicit("Pick a color", response_type=["red", "green", "blue"])  # enum
await ctx.elicit("Enter email", response_type=str)                         # string
await ctx.elicit("Confirm?", response_type=bool)                           # boolean

@dataclass
class ContactInfo:
    name: str
    email: str
await ctx.elicit("Contact details", response_type=ContactInfo)             # flat dataclass

Accepts: primitives, list[str] (becomes enum), dataclass, TypedDict, Pydantic BaseModel. All must be flat.

---

Security

MUST NOT request passwords, API keys, or tokens via elicitation — spec requirement. Those go through OAuth or user_config with sensitive: true (MCPB), not runtime forms.

---

When to escalate to widgets

Elicitation handles: confirm dialogs, enum pickers, short flat forms.

Reach for build-mcp-app widgets when you need:

• Nested or complex data structures
• Scrollable/searchable lists (100+ items)
• Visual preview before choosing (image thumbnails, file tree)
• Live-updating progress or streaming content
• Custom layouts, charts, maps

Remote Streamable-HTTP MCP Server — Scaffold

Minimal working servers in both recommended frameworks. Start here, then add tools.

---

TypeScript SDK (@modelcontextprotocol/sdk)

npm init -y
npm install @modelcontextprotocol/sdk zod express
npm install -D typescript @types/express @types/node tsx

`src/server.ts`

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

const server = new McpServer(
  { name: "my-service", version: "0.1.0" },
  { instructions: "Prefer search_items before calling get_item directly — IDs aren't guessable." },
);

// Pattern A: one tool per action
server.registerTool(
  "search_items",
  {
    description: "Search items by keyword. Returns up to `limit` matches ranked by relevance.",
    inputSchema: {
      query: z.string().describe("Search keywords"),
      limit: z.number().int().min(1).max(50).default(10),
    },
    annotations: { readOnlyHint: true },
  },
  async ({ query, limit }, extra) => {
    // extra.signal is an AbortSignal — check it in long loops for cancellation
    const results = await upstreamApi.search(query, limit);
    return {
      content: [{ type: "text", text: JSON.stringify(results, null, 2) }],
    };
  },
);

server.registerTool(
  "get_item",
  {
    description: "Fetch a single item by its ID.",
    inputSchema: { id: z.string() },
    annotations: { readOnlyHint: true },
  },
  async ({ id }) => {
    const item = await upstreamApi.get(id);
    return { content: [{ type: "text", text: JSON.stringify(item) }] };
  },
);

// Streamable HTTP transport (stateless mode — simplest)
const app = express();
app.use(express.json());

app.post("/mcp", async (req, res) => {
  const transport = new StreamableHTTPServerTransport({
    sessionIdGenerator: undefined, // stateless
  });
  res.on("close", () => transport.close());
  await server.connect(transport);
  await transport.handleRequest(req, res, req.body);
});

app.listen(process.env.PORT ?? 3000);

Stateless vs stateful: The snippet above creates a fresh transport per request (stateless). Fine for most API-wrapping servers. If tools need to share state across calls in a session (rare), use a session-keyed transport map — see the SDK's examples/server/simpleStreamableHttp.ts.

---

FastMCP 3.x (Python)

pip install fastmcp

`server.py`

from fastmcp import FastMCP

mcp = FastMCP(
    name="my-service",
    instructions="Prefer search_items before calling get_item directly — IDs aren't guessable.",
)

@mcp.tool(annotations={"readOnlyHint": True})
def search_items(query: str, limit: int = 10) -> list[dict]:
    """Search items by keyword. Returns up to `limit` matches ranked by relevance."""
    return upstream_api.search(query, limit)

@mcp.tool(annotations={"readOnlyHint": True})
def get_item(id: str) -> dict:
    """Fetch a single item by its ID."""
    return upstream_api.get(id)

if __name__ == "__main__":
    mcp.run(transport="http", host="0.0.0.0", port=3000)

FastMCP derives the JSON schema from type hints and the docstring becomes the tool description. Keep docstrings terse and action-oriented — they land in Claude's context window verbatim.

---

Search + execute pattern (large API surface)

When wrapping 50+ endpoints, don't register them all. Two tools:

const CATALOG = loadActionCatalog(); // { id, description, paramSchema }[]

server.registerTool(
  "search_actions",
  {
    description: "Find available actions matching an intent. Call this first to discover what's possible. Returns action IDs, descriptions, and parameter schemas.",
    inputSchema: { intent: z.string().describe("What you want to do, in plain English") },
    annotations: { readOnlyHint: true },
  },
  async ({ intent }) => {
    const matches = rankActions(CATALOG, intent).slice(0, 10);
    return { content: [{ type: "text", text: JSON.stringify(matches, null, 2) }] };
  },
);

server.registerTool(
  "execute_action",
  {
    description: "Execute an action by ID. Get the ID and params schema from search_actions first.",
    inputSchema: {
      action_id: z.string(),
      params: z.record(z.unknown()),
    },
  },
  async ({ action_id, params }) => {
    const action = CATALOG.find(a => a.id === action_id);
    if (!action) throw new Error(`Unknown action: ${action_id}`);
    validate(params, action.paramSchema);
    const result = await dispatch(action, params);
    return { content: [{ type: "text", text: JSON.stringify(result) }] };
  },
);

rankActions can be simple keyword matching to start. Upgrade to embeddings if precision matters.

---

Test it

The MCP Inspector connects to any transport and lets you poke tools interactively.

# Interactive — opens a UI on localhost:6274
npx @modelcontextprotocol/inspector
# → select "Streamable HTTP", paste http://localhost:3000/mcp, Connect

For scripted checks (CI, smoke tests):

npx @modelcontextprotocol/inspector --cli http://localhost:3000/mcp \
  --transport http --method tools/list

npx @modelcontextprotocol/inspector --cli http://localhost:3000/mcp \
  --transport http --method tools/call --tool-name search_items --tool-arg query=test

---

Connect users

Once deployed, users add the URL directly — no install step.

---

Deploy

Fastest path: Cloudflare Workers — two commands from zero to a live https:// URL on the free tier. Uses a Workers-native scaffold (not Express). → deploy-cloudflare-workers.md

This Express scaffold runs on any Node host — Render, Railway, Fly.io, a VPS. Containerize it (node:20-slim, copy, npm ci, node dist/server.js) and ship. FastMCP is the same story with a Python base image.

---

Deployment checklist

• [ ] POST /mcp responds to initialize with server capabilities
• [ ] tools/list returns your tools with complete schemas
• [ ] Errors return structured MCP errors, not HTTP 500s with HTML bodies
• [ ] CORS headers set if browser clients will connect
• [ ] Origin header validated on /mcp (spec MUST — DNS rebinding prevention)
• [ ] MCP-Protocol-Version header honored (return 400 for unsupported versions)
• [ ] instructions field set if tool-use needs hints
• [ ] Health check endpoint separate from /mcp (hosts poll it)
• [ ] Secrets from env vars, never hardcoded
• [ ] If OAuth: CIMD or DCR endpoint implemented — see auth.md

Resources & Prompts — the other two primitives

MCP defines three server-side primitives. Tools are model-controlled (Claude decides when to call them). The other two are different:

• Resources are application-controlled — the host decides what to pull into context
• Prompts are user-controlled — surfaced as slash commands or menu items

Most servers only need tools. Reach for these when the shape of your integration doesn't fit "Claude calls a function."

---

Resources

A resource is data identified by a URI. Unlike a tool, it's not called — it's read. The host browses available resources and decides which to load into context.

When a resource beats a tool:

• Large reference data (docs, schemas, configs) that Claude should be able to browse
• Content that changes independently of conversation (log files, live data)
• Anything where "Claude decides to fetch" is the wrong mental model

When a tool is better:

• The operation has side effects
• The result depends on parameters Claude chooses
• You want Claude (not the host UI) to decide when to pull it in

Static resources

// TypeScript SDK
server.registerResource(
  "config",
  "config://app/settings",
  { name: "App Settings", description: "Current configuration", mimeType: "application/json" },
  async (uri) => ({
    contents: [{ uri: uri.href, mimeType: "application/json", text: JSON.stringify(config) }],
  }),
);

# fastmcp
@mcp.resource("config://app/settings")
def get_settings() -> str:
    """Current application configuration."""
    return json.dumps(config)

Dynamic resources (URI templates)

RFC 6570 templates let one registration serve many URIs:

import { ResourceTemplate } from "@modelcontextprotocol/sdk/server/mcp.js";

server.registerResource(
  "file",
  new ResourceTemplate("file:///{path}", { list: undefined }),
  { name: "File", description: "Read a file from the workspace" },
  async (uri, { path }) => ({
    contents: [{ uri: uri.href, text: await fs.readFile(path, "utf8") }],
  }),
);

@mcp.resource("file:///{path}")
def read_file(path: str) -> str:
    return Path(path).read_text()

Subscriptions

Resources can notify the client when they change. Declare subscribe: true in capabilities, then emit notifications/resources/updated. The host re-reads. Useful for log tails, live dashboards, watched files.

---

Prompts

A prompt is a parameterized message template. The host surfaces it as a slash command or menu item. The user picks it, fills in arguments, and the resulting messages land in the conversation.

When to use: canned workflows users run repeatedly — /summarize-thread, /draft-reply, /explain-error. Near-zero code, high UX leverage.

server.registerPrompt(
  "summarize",
  {
    title: "Summarize document",
    description: "Generate a concise summary of the given text",
    argsSchema: { text: z.string(), max_words: z.string().optional() },
  },
  ({ text, max_words }) => ({
    messages: [{
      role: "user",
      content: { type: "text", text: `Summarize in ${max_words ?? "100"} words:\n\n${text}` },
    }],
  }),
);

@mcp.prompt
def summarize(text: str, max_words: str = "100") -> str:
    """Generate a concise summary of the given text."""
    return f"Summarize in {max_words} words:\n\n{text}"

Constraints:

• Arguments are string-only (no numbers, booleans, objects) — convert inside the handler
• Returns a messages[] array — can include embedded resources/images, not just text
• No side effects — the handler just builds a message, it doesn't do anything

---

Quick decision table

Server capabilities — the rest of the spec

Features beyond the three core primitives. Most are optional, a few are near-free wins.

---

instructions — system prompt injection

One line of config, lands directly in Claude's system prompt. Use it for tool-use hints that don't fit in individual tool descriptions.

const server = new McpServer(
  { name: "my-server", version: "1.0.0" },
  { instructions: "Always call search_items before get_item — IDs aren't guessable." },
);

mcp = FastMCP("my-server", instructions="Always call search_items before get_item — IDs aren't guessable.")

This is the highest-leverage one-liner in the spec. If Claude keeps misusing your tools, put the fix here.

---

Sampling — delegate LLM calls to the host

If your tool logic needs LLM inference (summarize, classify, generate), don't ship your own model client. Ask the host to do it.

// Inside a tool handler
const result = await extra.sendRequest({
  method: "sampling/createMessage",
  params: {
    messages: [{ role: "user", content: { type: "text", text: `Summarize: ${doc}` } }],
    maxTokens: 500,
  },
}, CreateMessageResultSchema);

# fastmcp
response = await ctx.sample("Summarize this document", context=doc)

Requires client support — check clientCapabilities.sampling first. Model preference hints are substring-matched ("claude-3-5" matches any Claude 3.5 variant).

---

Roots — query workspace boundaries

Instead of hardcoding a root directory, ask the host which directories the user approved.

const caps = server.getClientCapabilities();
if (caps?.roots) {
  const { roots } = await server.server.listRoots();
  // roots: [{ uri: "file:///home/user/project", name: "My Project" }]
}

roots = await ctx.list_roots()

Particularly relevant for MCPB local servers — see build-mcpb/references/local-security.md.

---

Logging — structured, level-aware

Better than stderr for remote servers. Client can filter by level.

// In a tool handler
await extra.sendNotification({
  method: "notifications/message",
  params: { level: "info", logger: "my-tool", data: { msg: "Processing", count: 42 } },
});

await ctx.info("Processing", count=42)   # also: ctx.debug, ctx.warning, ctx.error

Levels follow syslog: debug, info, notice, warning, error, critical, alert, emergency. Client sets minimum via logging/setLevel.

---

Progress — for long-running tools

Client sends a progressToken in request _meta. Server emits progress notifications against it.

async (args, extra) => {
  const token = extra._meta?.progressToken;
  for (let i = 0; i < 100; i++) {
    if (token !== undefined) {
      await extra.sendNotification({
        method: "notifications/progress",
        params: { progressToken: token, progress: i, total: 100, message: `Step ${i}` },
      });
    }
    await doStep(i);
  }
  return { content: [{ type: "text", text: "Done" }] };
}

async def long_task(ctx: Context) -> str:
    for i in range(100):
        await ctx.report_progress(progress=i, total=100, message=f"Step {i}")
        await do_step(i)
    return "Done"

---

Cancellation — honor the abort signal

Long tools should check the SDK-provided AbortSignal:

async (args, extra) => {
  for (const item of items) {
    if (extra.signal.aborted) throw new Error("Cancelled");
    await process(item);
  }
}

fastmcp handles this via asyncio cancellation — no explicit check needed if your handler is properly async.

---

Completion — autocomplete for prompt args

If you've registered prompts or resource templates with arguments, you can offer autocomplete:

server.registerPrompt("query", {
  argsSchema: {
    table: completable(z.string(), async (partial) => tables.filter(t => t.startsWith(partial))),
  },
}, ...);

Low priority unless your prompts have many valid values.

---

Which capabilities need client support?

Check client caps via server.getClientCapabilities() (TS) or ctx.session.client_params.capabilities (fastmcp) before using the bottom three.

Tool Design — Writing Tools Claude Uses Correctly

Tool schemas and descriptions are prompt engineering. They land directly in Claude's context and determine whether Claude picks the right tool with the right arguments. Most MCP integration bugs trace back to vague descriptions or loose schemas.

Anthropic Directory hard requirements

If this server will be submitted to the Anthropic Directory, the following are pass/fail review criteria (full list: https://claude.com/docs/connectors/building/review-criteria):

• Every tool must include readOnlyHint, destructiveHint, and title annotations — these determine auto-permissions in Claude.
• Tool names must be ≤64 characters.
• Read and write operations must be in separate tools. A single tool accepting both GET and POST/PUT/PATCH/DELETE is rejected — documenting safe vs unsafe within one tool's description does not satisfy this.
• Tool descriptions must not instruct Claude how to behave (e.g. "always do X", "you must call Y first", overriding system instructions, promoting products) — treated as prompt injection at review.
• Tools that accept freeform API endpoints/params must reference the target API's documentation in their description.

---

Descriptions

The description is the contract. It's the only thing Claude reads before deciding whether to call the tool. Write it like a one-line manpage entry plus disambiguating hints.

Good

search_issues — Search issues by keyword across title and body. Returns up
to `limit` results ranked by recency. Does NOT search comments or PRs —
use search_comments / search_prs for those.

• Says what it does
• Says what it returns
• Says what it doesn't do (prevents wrong-tool calls)

Bad

search_issues — Searches for issues.

Claude will call this for anything vaguely search-shaped, including things it can't do.

Disambiguate siblings

When two tools are similar, each description should say when to use the other one:

get_user      — Fetch a user by ID. If you only have an email, use find_user_by_email.
find_user_by_email — Look up a user by email address. Returns null if not found.

---

Parameter schemas

Tight schemas prevent bad calls. Every constraint you express in the schema is one fewer thing that can go wrong at runtime.

Describe every parameter. The .describe() text shows up in the schema Claude sees. Omitting it is leaving money on the table.

{
  query: z.string().describe("Keywords to search for. Supports quoted phrases."),
  status: z.enum(["open", "closed", "all"]).default("open")
    .describe("Filter by status. Use 'all' to include closed items."),
  limit: z.number().int().min(1).max(50).default(10)
    .describe("Max results. Hard cap at 50."),
}

---

Return shapes

Claude reads whatever you put in content[].text. Make it parseable.

Do:

• Return JSON for structured data (JSON.stringify(result, null, 2))
• Return short confirmations for mutations ("Created issue #123")
• Include IDs Claude will need for follow-up calls
• Truncate huge payloads and say so ("Showing 10 of 847 results. Refine the query to narrow down.")

Don't:

• Return raw HTML
• Return megabytes of unfiltered API response
• Return bare success with no identifier ("ok" after a create — Claude can't reference what it made)

---

How many tools?

The ceiling isn't a hard protocol limit — it's context-window economics. Every tool schema is tokens Claude spends every turn. Thirty tools with rich schemas can eat 3–5k tokens before the conversation even starts.

---

Errors

Return MCP tool errors, not exceptions that crash the transport. Include enough detail for Claude to recover or retry differently.

if (!item) {
  return {
    isError: true,
    content: [{
      type: "text",
      text: `Item ${id} not found. Use search_items to find valid IDs.`,
    }],
  };
}

The hint ("use search_items…") turns a dead end into a next step.

---

Tool annotations

Hints the host uses for UX — red confirm button for destructive, auto-approve for readonly. All default to unset (host assumes worst case).

server.registerTool("delete_file", {
  description: "Delete a file",
  inputSchema: { path: z.string() },
  annotations: { destructiveHint: true, idempotentHint: false },
}, handler);

@mcp.tool(annotations={"destructiveHint": True, "idempotentHint": False})
def delete_file(path: str) -> str:
    ...

Pair with the read/write split advice in build-mcpb/references/local-security.md — mark every read tool readOnlyHint: true.

---

Structured output

JSON.stringify(result) in a text block works, but the spec has first-class typed output: outputSchema + structuredContent. Clients can validate.

server.registerTool("get_weather", {
  description: "Get current weather",
  inputSchema: { city: z.string() },
  outputSchema: { temp: z.number(), conditions: z.string() },
}, async ({ city }) => {
  const data = await fetchWeather(city);
  return {
    content: [{ type: "text", text: JSON.stringify(data) }],  // backward compat
    structuredContent: data,                                    // typed output
  };
});

Always include the text fallback — not all hosts read structuredContent yet.

---

Content types beyond text

Tools can return more than strings:

`resource_link` vs embedded: link for large payloads or when the client might not need it (let them decide). Embed when it's small and always needed.

Version pins

Every version-sensitive claim in this skill, in one place. When updating the skill, check these first.

How to verify

# ext-apps latest
npm view @modelcontextprotocol/ext-apps version

# CF template still exists
gh api repos/cloudflare/ai/contents/demos/remote-mcp-authless/src/index.ts --jq '.sha'

# MCPB schema
curl -sI https://raw.githubusercontent.com/anthropics/mcpb/main/schemas/mcpb-manifest-v0.4.schema.json | head -1