Lesson 17.1117.11 How today's tools realize it

In everyday life. Every outlet in a country follows one standard, so any lamp works in any room, though some rooms add a dimmer or a timer.

What you will learn. How Claude Code, Codex and OpenCode each connect to an MCP server, where each adds to the specification or differs from it, and where the servers run.

Why it matters. The server stays the same, but the settings file, the way a key travels and who approves a call change with each product.

Concepts 17.1 to 17.10 taught the standard. This concept shows how each product connects to an MCP server, as of November 1, 2026, against revision 2026-07-28.¹ The field guide that follows has the commands.

Which products does this page cover?

You know three from Chapter 12: Claude Code, Codex and OpenCode, coding agents that act as MCP hosts. Three more serve programs. The Messages API is how a program sends a conversation to Claude and gets a reply. Its MCP connector, Anthropic's name for this one feature, lets the request name a remote MCP server for Claude to call. The Responses API is OpenAI's equivalent, and can call MCP servers too. The Claude Agent SDK is a library for building your own agent around Claude Code's loop. The lab does not use them, but the same server serves them too.

How far to trust this page. Each statement below is what the product's own documentation said on November 1, 2026, cited to its source. Brightline's servers have been run against the MCP Inspector and the official MCP code libraries, and the TypeScript server against Claude Code. They have not yet been run against Codex or OpenCode, so note what you see in the lab.

How does each one connect, and where does it differ?

Pick the coding agent you use. Each tab also covers that vendor's programming interface, and each bullet says what the product adds to the MCP specification or how it sets up a connector.

Where do the servers run?

Cloudflare Workers is the book's platform for stateless servers. A Worker is a small program Cloudflare runs at a web address, on the machine nearest each caller, in a light, short-lived sandbox: what 17.4 and 17.8 ask of a stateless server. Wrangler, Cloudflare's command-line tool, runs it locally (wrangler dev, no account needed) and uploads it (wrangler deploy). The code exports a fetch function that receives each request and an env object holding its bindings: its secrets and handles to platform resources, such as env.ASSETS for its files. Brightline's TypeScript server is a standard fetch function built with the official MCP SDK, so the platform stays a deployment choice. Never heard of Cloudflare? Field guide F1 starts with "Meet Cloudflare": what the company does, why it matters for AI, and what its free plan gives you, then explains each term from zero.

• Staging and production. One configuration file defines production and staging, each a separate Worker with its own address and secrets. Bindings are not inherited, so each lists its own.¹⁰
• Each deploy can be undone. Every deploy is kept as a numbered Worker version. A rollback (wrangler rollback) makes one of the 100 most recent live again.¹¹ Version URLs give a version a public address before it goes live, and are off for Brightline's production Worker.¹²
• What gets logged. Workers Logs keep console output and errors, on by default for new Workers.¹³ wrangler tail shows live requests.¹⁴
• Python. Python Workers run on Pyodide, Python compiled to WebAssembly, a compact program format browsers can run.¹⁵ So only packages written purely in Python, or built for WebAssembly, run.¹⁶
• Security. The address is public. The Worker checks Origin and the key on every request, and never serves the data files directly. The CI deploy credential is limited to Workers and what they use, in one account.¹⁷
• Status and price. Generally available. The free plan gives 100,000 requests a day and 10 ms of processor time (CPU time) per request. The paid plan allows 30 seconds by default.¹⁸ The free plan keeps Workers Logs 3 days, until new log pricing starts on December 1, 2026.¹⁹ Cloudflare's own Containers are not on the free plan.²⁰

The commands are in field guide F6. Another hosting service can run the same server, with only the configuration file and deploy commands changed.

The differences that change a decision

1. The server never changes. Brightline's servers connected unchanged to the official SDK clients and the MCP Inspector, and the TypeScript server to Claude Code. Hosts write three facts in their own files and syntax: the address, the key, and which tools are allowed.
2. The revision is set differently. Claude Code uses 2026-07-28 with servers that support it. OpenCode needs "protocol": "2026-07-28". Codex's page states none: record what your run shows.
3. Read-only marks are trusted differently. Codex can skip approval for tools a server marks read-only: safe for a server you wrote, dangerous for a stranger's.
4. A key travels differently: ${VAR} in Claude Code, bearer_token_env_var in Codex, {env:NAME} in OpenCode, never in the file. The Agent Plugins core, the plugin standard from Chapter 16, cannot carry a key at all, so a shared-key server is set per host, and OAuth lets one entry serve every host.
5. Records of each call are uneven. Claude Code and Codex export their own records. OpenCode's documentation describes none. Lab Task 16c records whether any sends its trace ID (traceparent) to the server.²¹

What stays the same on every product?

1. The server is written once, to the open protocol. Whichever host connects, it validates its inputs, checks every request's credential and Origin, and returns only what the caller needs.
2. The host decides whether a call may run. A server's own description of itself never decides it alone.
3. A span shows that a call ran. Only a test or an acceptance check shows it was right.

For the exam. CCDV-F and CCAR-F ask about Claude's behavior. Where Claude Code differs from the specification, such as its output limit or its tool naming in permission rules, the answer that scores is Claude Code's behavior.