MCP has a neutral home now: decide your server policy before your users do
· By Bee Flow · 5 min read
MCP moved to the new Agentic AI Foundation on 9 December, two weeks after a specification revision that reworked authorization. Neither decides which MCP servers your organisation allows. What an administrator should settle first.

On 9 December 2025 the Linux Foundation announced the Agentic AI Foundation, a new home for open standards around AI agents. Anthropic contributed the Model Context Protocol, next to Block's goose and OpenAI's AGENTS.md. Two weeks earlier, on 25 November, MCP turned one and shipped a new specification revision, dated 2025-11-25, with many of its changes in authorization.
If you run the systems an AI assistant will connect to, those two items belong together. The protocol is now less tied to one company, and the rules for how a client gets access to a server are more concrete. Neither tells you which servers your organisation should allow. That part is yours, and it is easier to settle before people start adding servers on their own.
What the move to a foundation changes
Less than the headlines suggest, by the project's own account. The MCP announcement says the governance model introduced earlier this year "continues as is", that the maintainers keep making the decisions with community input through the SEP process, and that the foundation "will not dictate the technical direction of MCP". The Linux Foundation release lists AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as platinum members.
For an administrator the practical meaning is stability. A protocol with a neutral steward and several large vendors behind it is a reasonable thing to write internal policy against.
It does not mean anyone vets the servers. The official MCP Registry, in preview since 8 September 2025, describes itself as an open catalogue of self-reported information, with moderation that relies on community reports and a denylist. The same post expects companies with strict privacy and security requirements to run private subregistries on top of it. We read that as a hint: a list of your own is expected.
What the November revision says about authorization
Authorization in MCP is optional. Where a server speaks HTTP and uses it, the revision now recommends OAuth Client ID Metadata Documents: the client's identity is an https URL pointing at a JSON document the client hosts. Dynamic Client Registration goes from "SHOULD" in the June 2025 revision to "MAY", kept for backwards compatibility. For local servers over stdio the spec says implementations should not use this flow and should take credentials from the environment.
A few details matter for whoever runs the identity provider. Authorization servers MAY apply domain-based trust policies, anywhere from an allowlist of trusted domains to accepting any https client. The spec also warns that a metadata document cannot stop someone from posing as a known client through a localhost redirect, and asks authorization servers to show the redirect host and warn about localhost-only redirects.
Scopes are meant to grow during use. A server can answer with insufficient_scope and the scopes it needs, after which the client asks the user to approve more (the step-up flow). The security best practices name the mistakes: wildcard scopes such as * or full-access, and publishing every scope up front. Tokens must be issued for the specific server, and a server may not pass a client's token on to another API.
Two more changes touch trust directly. URL mode elicitation lets a server send the user to a web page for things like an API key or a third-party login, so the secret never passes through the client. The client must show the full URL and ask for consent before opening it. And for clients that offer one-click setup of local servers, the best practices require a consent dialog with the exact command that will run. URL mode is new in this revision and the spec says its design may still change. The same goes for the experimental tasks feature.

Decisions to make before users add servers
The spec describes what clients and servers must do. Your organisation's choices it leaves to you. These are the ones we would settle first.
Who may add a server. One model is that only administrators install servers and users pick from what is there. Another is that users may add servers, but only from a list. A consent dialog with the exact command is a good last check for one person on one laptop. It is not a policy for a team.
Remote or local. A local server is a program that runs with the client's privileges, on someone's machine or on your server. A remote server is someone else's service that sees whatever its tools receive. Both have their place. Write down which kinds you accept, and where local ones may run.
Which hosts. Keep your own list of approved servers and hosts, with an owner for each entry. A registry listing tells you a server exists and who says they publish it. It says nothing about what happens to your data there.
Whose credentials. With a shared key every user acts with that key's permissions, and the server's logs show one identity. Per-user OAuth keeps the audit trail personal and lets you revoke one person. For systems holding personal or financial data we would want per-user access, or a shared key with read-only rights at most.
Which scopes. Start with read access. Expect step-up prompts in the middle of a task once write access is needed, and tell people which requests to approve and which to refuse or report.
Your identity provider. If Keycloak, Entra ID or something similar will be the authorization server for your MCP servers, check whether it supports Client ID Metadata Documents yet. If it does, decide whether it accepts any https client_id or only clients on a list, and look at how its consent screen shows a localhost redirect.
Leaving and logging. Know where tool calls are logged, and how you revoke a person's tokens on every connected server when they leave or change roles.
None of this waits for clients and servers to adopt the 2025-11-25 revision. The project calls the release backward compatible, so expect a mix of old and new behaviour for a while, and ask each vendor which revision they implement.
How Bee Flow fits in
Bee Flow is a shared AI workspace in which administrators decide which outside systems the team's assistant may reach, instead of each user wiring up tools alone. You can run it on your own server, so the connections to your business systems start from a machine you control. The integrations page shows what it connects to, and self-host or EU cloud explains the hosting options. If you are drafting an MCP policy and want to see how it maps onto a team workspace, get in touch.