Technology & AI 20 November 2026 7 min read

MCP and A2A: Connecting AI Agents Without Opening Every Door

Engineer at a laptop with a glowing gear and connection lines on the screens behind

Most enterprises that have tried to put AI agents to work hit the same wall. The model is capable enough, but every system it needs to touch (the CRM, the ticketing tool, the document store, the finance platform) needs its own connector. Build five agents against ten systems and you can end up maintaining something close to fifty bespoke integrations, each with its own credentials, quirks and ways of failing.

Two open protocols are now the main attempt to fix that. The Model Context Protocol (MCP) standardises how an agent connects to tools and data. The Agent2Agent protocol (A2A) standardises how one agent talks to another. Both now sit under neutral governance and are turning up in products your teams already buy.

That is good news for integration cost. It is less good news if anyone in the organisation assumes a standard protocol is the same thing as a secure one. The MCP specification itself is clear that it is not.

What MCP and A2A actually do

MCP came first. Anthropic introduced it in late 2024 as an open way to connect language model applications to outside systems. An AI application (the host) connects to servers that expose resources, prompts and tools. The practical effect is that a system owner can publish one MCP server and any compliant AI application can use it.

A2A addresses a different problem. Google announced it in April 2025 so that agents built on different platforms could find each other, describe what they can do and pass work back and forth. Each agent publishes an Agent Card, a small JSON document setting out its skills, where to reach it and how to authenticate.

A simple way to hold the distinction: MCP is how your procurement agent reads the contract repository. A2A is how your procurement agent asks a supplier's sales agent for a revised quote.

Who governs them, and how mature they are

Governance matters because nobody wants to build integration architecture on a protocol one vendor can change at will.

  • MCP. In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. At the time Anthropic reported more than 10,000 active public MCP servers and adoption in ChatGPT, Gemini, Microsoft Copilot, Cursor and Visual Studio Code.
  • A2A. Google moved A2A to the Linux Foundation in June 2025, citing support from more than 100 technology companies. The specification reached version 1.0 in March 2026, and in August 2026 the project joined the same Agentic AI Foundation.

Neither standard is frozen. MCP's current revision, dated 28 July 2026, changed some fundamentals: requests are now stateless and dynamic client registration is deprecated. A2A's 1.0 guidance includes a migration section for earlier adopters. Plan for upgrades as you would with any young API standard, and keep business logic away from protocol details.

Why standards lower integration cost

The economics will be familiar to anyone who has run an API-first integration programme. Point-to-point connections multiply; standard interfaces add up. If each system of record exposes one well-governed MCP server, every new agent reuses it instead of commissioning another connector. If your agents and your vendors' agents both speak A2A, a cross-company workflow becomes a configuration and contract question rather than a build.

A standard also gives security and operations teams one pattern to review and monitor, instead of a homegrown mechanism per project. That only pays off if you insist on it. Left alone, teams will stand up MCP servers the way they once stood up shadow APIs: quickly, locally and without telling anyone.

A shared protocol makes it cheap to connect things. Deciding which things should be connected is still your job.

What the protocols leave to you

The MCP specification is candid about this. It says MCP "cannot enforce these security principles at the protocol level" and puts the burden on implementers. Five areas stay firmly with you.

  • Authorisation and delegation. When an agent calls a tool on behalf of a junior analyst in accounts payable, whose permissions apply: the analyst's, the agent's, or the service account someone created in a hurry? MCP's authorisation model for remote servers builds on OAuth 2.1, and its guidance forbids passing through tokens issued for something else. That is plumbing. Deciding that the agent must never inherit a finance director's rights is a policy decision you make and enforce.
  • Tool permissions. The specification treats tools as arbitrary code execution and says tool descriptions should be considered untrusted unless they come from a trusted server. Its guidance warns against wildcard scopes and recommends starting with minimal permissions and adding more only when needed. In practice that means an allow-list of approved servers and tools for each agent, not "connect whatever looks useful".
  • Prompt injection through tool outputs. An agent that reads an email, a web page or a supplier document can be steered by instructions hidden in that content. The OWASP Top 10 for Agentic Applications, published in December 2025, ranks agent goal hijack first and tool misuse second. No protocol can fix this, because the attack travels inside legitimate data. The defences are architectural: treat tool output as untrusted input, separate reading from acting, and require human approval for consequential actions.
  • Audit logging. You need to reconstruct who asked for what, which agent acted, which tools it called with which arguments, and what came back. Protocols carry the messages; they do not keep your records. Our piece on agentic AI from pilots to production covers the decision logging auditors will expect.
  • Supply chain. A third-party MCP server is software running with your agent's privileges, and the specification's guidance describes how a local server installed without consent checks or sandboxing can run malicious commands. Vet servers like any other dependency.

A gateway pattern that holds up

Most organisations will end up with a gateway between their agents and the servers and peers they use. That is the right instinct, provided it is designed as a control point and not simply a proxy.

A sensible gateway does four jobs. It holds the registry of approved servers, tools and external agents, so nothing is connected that has not been reviewed. It enforces identity and scope, issuing short-lived, narrowly scoped tokens per task instead of letting agents carry standing credentials. It logs every call and response in a format your security team already monitors. And it applies policy, such as blocking writes to the general ledger outside approved workflows or sending payments above a threshold to a person.

Picture a hypothetical insurer whose claims agent calls a policy lookup and a payments system. Without a gateway, both connections share one broad service credential and the logs live in two places. With one, the payments call is scoped to a single claim, capped in value, logged centrally and paused for approval when the amount looks unusual. Same protocol, very different risk.

Gateways add latency and another component to run, and a badly governed one becomes a single point of failure. Keep it thin and include it in your AI red-teaming.

Common questions

What is the difference between MCP and A2A?

MCP connects an AI agent to tools and data sources, such as a database, a document store or a business application. A2A lets independent agents, possibly built by different vendors, discover each other and exchange tasks. Many enterprise designs will use both: MCP for an agent's own toolset, A2A between agents.

Is the Model Context Protocol secure?

MCP includes an OAuth-based authorisation framework for remote servers and detailed security guidance, but the specification states that it cannot enforce security principles at the protocol level. Whether a deployment is secure depends on how you handle permissions, consent, logging and untrusted tool output.

Who owns MCP and A2A now?

Both are open source projects under the Agentic AI Foundation, which sits within the Linux Foundation. Anthropic donated MCP in December 2025. A2A, created by Google, moved to the Linux Foundation in June 2025 and joined the Agentic AI Foundation in August 2026.

Where Ganexa can help

Ganexa helps technology leaders design the integration and control architecture that lets agents use enterprise systems without inheriting more access than they need, including server registries, gateway policy, delegated identity and audit trails. Our cybersecurity for the AI era team works alongside your architects to make agent connections reviewable before they scale. If you are about to connect your first agents to production systems, book a free consultation.

AI AgentsMCPA2AIntegrationAI Security