Skip to content
Blentera

Resources, AI agents

The Model Context Protocol (MCP) explained

The Model Context Protocol (MCP) is an open standard that connects AI applications and agents to tools and data. This guide explains how it works, who maintains it today, where the security risks lie and how to decide which MCP servers your agents may use.

Updated 20 November 2026. 10 minute read.

In short

  • MCP is an open protocol, introduced by Anthropic in November 2024, that gives AI applications one standard way to connect to tools, data and prompts offered by MCP servers.
  • Since 9 December 2025, MCP has been a project of the Agentic AI Foundation, a directed fund under the Linux Foundation. Its maintainers still decide the technical direction. The current specification is version 2026-07-28.
  • An AI application (the host) runs one MCP client per server. Servers offer tools, resources and prompts. Local servers run on the same machine, usually over stdio. Remote servers run over HTTP and use OAuth-based authorisation.
  • The main risks are tool poisoning, over-broad permissions and prompt injection through tool output. The specification treats tools as arbitrary code execution and says hosts must obtain consent before a tool is used.
  • Govern MCP with an approved list of servers, least-privilege scopes per agent, approval for actions that change data or leave the company, and a log of every tool call.

What MCP is

The Model Context Protocol is an open protocol that lets AI applications connect to external data sources and tools in a standard way. Anthropic open-sourced it on 25 November 2024. Before MCP, every pairing of an AI application with a business system needed its own integration. With MCP, a system is wrapped once as an MCP server, and any compatible AI application can use it.

The specification says MCP takes inspiration from the Language Server Protocol, which standardised how code editors support programming languages. MCP does the same for context and tools in AI applications. By December 2025, ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code all supported it as clients. Official software development kits in TypeScript, Python, C#, Go, Rust and Ruby now meet the project's top support tier.

Who maintains MCP today

On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation. Anthropic, Block and OpenAI co-founded the foundation, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. MCP was a founding project alongside Block's goose and OpenAI's AGENTS.md.

The AAIF governing board decides on strategic investment, budget, members and new projects. Each project keeps full autonomy over its technical direction. For MCP, decisions about the protocol stay with its maintainers, guided by public Specification Enhancement Proposals (SEPs) and working groups. Code and specifications are licensed under Apache 2.0.

The current specification is version 2026-07-28, released on 28 July 2026. It made the protocol stateless: the opening handshake and protocol-level sessions are gone, and each request carries its own version and capabilities. It also added a formal extensions framework and a deprecation policy with a minimum window of twelve months. The roadmap published in August 2026 makes agent identity and enterprise-ready security one of five priority areas.

Hosts, clients, servers and tools

MCP uses JSON-RPC 2.0 messages between three roles:

  • Host: the AI application or agent platform that the employee or agent works in.
  • Client: a connector inside the host. The host runs one client for each server it connects to, which keeps servers isolated from each other.
  • Server: a program that offers capabilities from a system, such as a CRM, a file store or a code repository.

Servers offer three kinds of feature. Tools are functions the model can call, such as searching tickets or creating an invoice. Resources are data the model or the employee can read. Prompts are templates for common tasks. Clients can offer elicitation, so a server can ask for more information during a task. In the 2026-07-28 version, a server does this by replying that input is required, and the client retries the request with the answer.

A typical tool call works like this. The host gets the list of tools and their descriptions from each server. The model reads those descriptions and chooses a tool. The client sends the call, the server runs it and returns a result, and the result goes back into the model's context. Every step of that chain is text the model reads, which matters for security.

Optional extensions add features on top of the core protocol. Examples include Tasks for long-running work, MCP Apps for interactive elements inside a conversation, and Enterprise-Managed Authorization.

Remote versus local servers

MCP defines two standard transports. They lead to different risks.

  • Local servers use stdio: the client starts the server as a program on the same machine and talks to it through standard input and output. The server runs with the same privileges as the client. The MCP security guidance notes that the stdio transport is not a sandbox, and a local server can reach whatever the machine can reach.
  • Remote servers use Streamable HTTP: the server runs as a web service, often operated by the vendor of the system it connects to. The 2026-07-28 version makes a remote server work like any other HTTP service, so it can sit behind standard load balancers and gateways.

For a company, remote servers are usually easier to govern. Access runs through authorisation that can be granted and revoked centrally, and nothing new is installed on employee machines. The MCP guidance on local servers suggests asking first whether a server needs to run locally at all: reading local files or driving a local tool are good reasons, but wrapping a remote API is not.

Where a local server is needed, the MCP security guidance requires that one-click setup shows the exact command before it runs and asks for explicit approval. It recommends running local servers in a sandbox with minimal privileges. The specification requires HTTP servers to validate the Origin header, and local ones should listen only on localhost.

Authentication and authorisation

Authorisation is optional in MCP, but a remote server that holds company data should always use it. For HTTP servers, the specification builds on OAuth 2.1. The MCP server acts as an OAuth resource server, the MCP client as an OAuth client, and a separate authorisation server issues access tokens. Local stdio servers should not use this flow and get credentials from their environment instead.

  • Discovery: servers must publish OAuth Protected Resource Metadata (RFC 9728), so clients can find the right authorisation server.
  • Client registration: Client ID Metadata Documents are the preferred method. Dynamic Client Registration was deprecated in version 2026-07-28 and is kept only for compatibility.
  • Token audience: a server must accept only tokens issued for itself. Passing a client's token straight through to another API is explicitly forbidden.
  • Issuer checks: clients must validate the issuer in the authorisation response (RFC 9207) before redeeming a code, which closes a known mix-up attack.
  • Scopes: the security guidance recommends starting with a minimal scope and asking for more only when a privileged operation is first attempted. It warns against wildcard scopes such as full access.

The Enterprise-Managed Authorization extension, declared stable in June 2026, lets an organisation provision MCP server access centrally through its identity provider. Employees get the servers approved for their groups and roles at sign-in, instead of approving each server one by one.

Security risks

The specification states that tools represent arbitrary code execution and must be treated with caution. Its security principles say hosts must obtain explicit consent before invoking any tool. Three risks matter most for companies.

Tool poisoning

Tool names and descriptions are instructions the model reads. In April 2025, Invariant Labs showed that a server can hide directives in a tool description that the employee never sees but the model follows. The MCP security guidance describes related attacks: tool shadowing, where one server's description changes how the agent uses another server's tools, and rug pulls, where a server changes its tool definitions after approval. A one-time review at installation does not cover these.

Over-broad permissions

An agent connected with an all-access token can do everything that token allows. The MCP guidance lists the effects of broad scopes: a larger blast radius if a token leaks, harder revocation and less useful audit trails. OWASP calls the wider problem excessive agency. Tool annotations such as read-only or destructive are hints only, and the specification says clients must treat them as untrusted unless they come from a trusted server.

Prompt injection through tool output

Results returned by tools, such as an email, a web page or a support ticket, can contain instructions written by an attacker. Simon Willison's lethal trifecta names the dangerous combination: access to private data, exposure to untrusted content and the ability to communicate externally. MCP makes it easy to assemble all three from different servers in one session. The specification asks servers to sanitise tool outputs and clients to validate results before passing them to the model, but neither fully prevents injection.

The risk belongs to the combination of servers an agent can use, not to one server alone. Review the full set of tools each agent has.

How to govern which MCP servers agents may use

  1. Keep an approved list. Allow agents to connect only to MCP servers your company has reviewed. The official MCP Registry, launched in preview in September 2025, is designed so that companies can build private sub-registries on top of it.
  2. Check provenance. Prefer servers run by the vendor of the system itself. For open-source servers, check the publisher, the repository and the maintenance history, and pin a version.
  3. Watch for changes. Record each server's tool definitions at approval and review any change before agents use it.
  4. Grant least privilege per agent. Connect each agent only to the servers and tools its role needs, with the narrowest scopes available. Prefer read-only access where it is enough.
  5. Require approval for risky tools. Actions that change data, spend money, delete anything or send anything outside the company should need an employee's approval.
  6. Break the trifecta. Avoid giving one agent private data, untrusted content and an outbound channel at once. If it must have all three, require approval for outbound actions.
  7. Log every tool call. Record the agent, server, tool, inputs, result and approval, and review the log regularly.
  8. Revoke quickly. Be able to disconnect a server or revoke an agent's token at once, and remove access when an agent's role changes.

How Blentera helps

Blentera is a platform where companies run AI agents with employees in charge. Agents reach apps and MCP servers through connectors. Approval rules can be set per agent, project, app or tool, so an MCP tool that changes data can require an employee's approval. Each agent has a role in the org chart, budgets warn at 80% and pause at the limit, and an activity log records what agents do. Blentera is hosted in the EU.

Questions

Sources

  1. Model Context Protocol specification, version 2026-07-28
  2. MCP specification 2026-07-28: key changes
  3. MCP specification 2026-07-28: architecture
  4. MCP specification 2026-07-28: transports
  5. MCP specification 2026-07-28: authorization
  6. MCP specification 2026-07-28: tools
  7. MCP security best practices
  8. MCP local server security
  9. MCP governance and licensing (GOVERNANCE.md), GitHub
  10. Introducing the Model Context Protocol, Anthropic, 25 November 2024
  11. Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic, 9 December 2025
  12. MCP joins the Agentic AI Foundation, MCP blog, 9 December 2025
  13. The 2026-07-28 specification, MCP blog, 28 July 2026
  14. The new MCP roadmap, MCP blog, 22 August 2026
  15. Enterprise-Managed Authorization: zero-touch OAuth for MCP, MCP blog, 18 June 2026
  16. Introducing the MCP Registry, MCP blog, 8 September 2025
  17. The official Ruby SDK for MCP reaches Tier 1, MCP blog, 30 October 2026
  18. Tool annotations as risk vocabulary, MCP blog, 16 March 2026
  19. MCP security notification: tool poisoning attacks, Invariant Labs, 1 April 2025
  20. The lethal trifecta for AI agents, Simon Willison, 16 June 2025
  21. OWASP Top 10 for LLM Applications 2025: LLM06 Excessive Agency, OWASP GenAI Security Project

This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.

Get started

Put your first agent to work in minutes.

Three steps, and your first agent is working for your company.

  1. Step 1

    Create an account

    Sign up with your work email and invite your team. No sales call needed.

  2. Step 2

    Add your first agent

    Create an agent, give it a role, and connect your own AI key and the apps it may use.

  3. Step 3

    Give it a task

    Ask in Blentera or your chat app, and follow the work until it is done.