RXed AI News

What it is — in one coffee-break

Protocols are the plumbing standards that let an AI agent reach your tools, and other agents, without a custom integration for every pair. Two matter today. MCP (Model Context Protocol) is agent-to-tool: you expose your booking system, invoice database or CRM once, as an MCP server, and any client that speaks MCP — Claude, Copilot Studio, Cursor, a home-grown agent — can use it. A2A (Agent2Agent) is agent-to-agent: your supplier's agent and yours exchange tasks while each keeps its own memory and secrets.

This is the anti-lock-in element. Without a protocol, N tools times M assistants means N×M connectors, each owned by a vendor who can deprecate it. With one, it is N+M. Write the server once; swap the model underneath when a cheaper one lands. That is the whole argument.

Both standards are now governed outside their founders. The current MCP revision, 2026-07-28, drops the stateful handshake so servers sit behind ordinary load balancers, and formally deprecates Dynamic Client Registration. Google donated A2A to the Linux Foundation, where it passed 150+ backing organisations and v1.0. A third layer is arriving: agent payments — ACP (OpenAI + Stripe) for conversational checkout, Google's AP2 for cryptographically signed spending mandates, Stripe/Tempo's MPP for pre-authorised micropayment sessions, and Coinbase's x402 for stablecoin settlement over HTTP 402.

When you actually need it (and when you don't)

You need it when an agent must reach data that is not in a document — live stock, open jobs, customer balances, tomorrow's route. Also when you already run two or more AI tools and refuse to pay for the same integration twice, or when a vendor's roadmap is the only thing between you and your own data.

You do not need it if the job is "summarise this PDF" or "draft this email" — a chat window with file upload is protocol-free and finished. Same if you have one system of record that already ships a native, supported integration with the assistant you use; a bespoke server buys you a maintenance job. And if nobody on your team reads log files, skip it. An MCP server is running code holding credentials, not a setting you toggle.

How to recognize good vs bad implementations

Good looks like: a documented tool list you can read, per-tool scopes, OAuth with real issuer validation, read-only variants of write tools, and an audit log. Bad looks like: one tool called execute, a single admin token, no logging. The NSA's May 2026 guidance is blunt — MCP's spread "has outpaced the development of its security model" — and the record backs it: a proxy package with a 9.6-severity remote-code-execution flaw downloaded over 437,000 times, a poisoned GitHub issue that walked an agent into private repositories, a fake Postmark server that silently BCC'd every email to the attacker.

Three questions that end a bad demo fast. Show me the exact list of tools this server exposes, and which of them can write or delete. Which credential does the server run as — mine, or a service account holding everyone's data? If a customer pastes hostile text into a support ticket, walk me through what stops the agent acting on it. Vendors who have thought about it answer in thirty seconds. The rest pivot to model quality.

What this costs

The protocol itself is free: open specs, permissively licensed SDKs in TypeScript, Python, Go, C# and Rust. Writing your own MCP server costs developer time — a read-only server over an existing REST API or Postgres table is roughly half a day to two days, about €400–€1,600 at €100/hour, and it then works with every MCP client. Hosting is a small VM or a Cloudflare Worker at €5–€20 a month. The real spend sits elsewhere: client seats at €20–€40 per user per month, plus tokens. Commercial MCP gateways and registries start near €10–€25 per user per month; under about 50 people, a config-file allowlist and a scoped service account do the same work for zero. On payments, x402 charges no protocol fee — you pay chain gas only — while ACP and MPP settle through Stripe at roughly 2.9% + €0.25 per transaction. The protocol is free; moving the money is not.

Where to see it scored

We scored 59 of 62 audited tools on Pc. Top of the table: Glean (9.5), Gemini / Google (9), Claude / Anthropic (9) and Windsurf, Devin Desktop (8.5) — all of them ship both an MCP server and an MCP client, so they compose with tools you already own. Bottom: Suno (1.5) and Midjourney (1), creative generators with no protocol surface at all — no server, no client, no agent card. You drive them by hand or through a proprietary API.