# What is Model Context Protocol?

> The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data. An AI app's client calls MCP servers that offer tools; protected servers accept only OAuth tokens issued for them.

Explainer by Boston Identity, 2026-10-15. 7 min read. https://insights.bostonidentity.com/glossary/mcp

- What it is: protocol
- Defined by: MCP project specification, revision 2026-07-28
- Who it's for: Teams connecting AI assistants and agents to company tools and data

## What is Model Context Protocol?

MCP is an open protocol for connecting applications built on large language models (LLMs) to outside data sources and tools [S1]. An AI coding tool, a chat app or a custom AI workflow can use it as one standard way to give a model the context it needs [S1].

The application the user works in is the host, and it starts the connections [S1]. Inside the host, a client is the connector that talks to an MCP server, a service that offers context and capabilities; the messages between them are JSON-RPC 2.0 [S1]. A server can offer resources (data), prompts (message templates) and tools, which are functions the model can run [S1].

Authorization is optional in MCP [S4]. When a server on an HTTP connection uses it, the server acts as an OAuth 2.1 resource server, the service that accepts access tokens, and the client acts as an OAuth 2.1 client on the user's behalf [S4 §Roles]. The specification does not cover how the authorization server is built [S4 §Roles]. It does require that server, which signs the user in and issues tokens, to implement OAuth 2.1 and to publish discovery metadata under RFC 8414 or OpenID Connect Discovery [S4 §Overview]. An implementation that connects over stdio (standard input and output) should not use this flow, and gets its credentials from the environment instead [S4].

Think of it like this: MCP works like a standard wall socket for AI applications: any app with the right plug can draw on any server built to the standard. The comparison stops at safety. A socket never asks who is plugging in, while a protected MCP server checks an access token on every request [S4 §Token Requirements].

Not the same as: MCP is not a policy system. It says how a client gets a client ID and an OAuth token and how it presents the token [S4]. The decision about who may do what is moving up into identity providers and agent gateways [BI-FIND-2026-0069](https://insights.bostonidentity.com/ref/BI-FIND-2026-0069).

## How does Model Context Protocol work?

We follow one host application, with its MCP client, as it reaches a protected MCP server over HTTP and gets a token from that server's authorization server. The steps use the current revision, in which every request stands on its own [S2].

[Figure: The MCP client checks the version, gets a token for one server from that server's authorization server, then lists and calls tools, and asks for more scope only when the server names what is missing.]

1. **The client checks the protocol version.** Every request names the protocol version it uses, in its _meta field and, on HTTP, in the MCP-Protocol-Version header [S2]. A server that does not support that version must answer with an error that lists the versions it does support, and the client should retry with one they share [S2]. Servers must offer a server/discover call, which a client may send first to learn those versions [S2]. Revisions up to 2025-11-25 opened with an initialize handshake; the current revision has none [S2].
2. **The server points to its authorization server.** A protected server must answer a request with a missing, invalid or expired token with HTTP 401 [S4 §Error Handling]. Its WWW-Authenticate header should name the scopes the request needs [S4 §Scope Selection Strategy]. The server must publish Protected Resource Metadata (RFC 9728), a document that tells clients which authorization server it uses, and clients must use it to find that server [S4 §Overview]. The client then reads the authorization server's own metadata, published under RFC 8414 or OpenID Connect Discovery; clients must support both [S4 §Overview].
3. **The client registers and the user signs in.** Before the flow starts, the client must get a client ID [S4 §Client Registration]. Clients and authorization servers should support Client ID Metadata Documents, in which a client names itself with a web address that points to its own description [S4 §Overview] [BI-FIND-2026-0704](https://insights.bostonidentity.com/ref/BI-FIND-2026-0704). Pre-registration is another route. Dynamic Client Registration is deprecated and kept only for authorization servers that do not support metadata documents [S4]. The authorization server signs the user in if needed and issues the token [S4 §Roles]. The client must name the MCP server in a resource parameter on both the authorization and token requests [S4 §Resource Parameter Implementation]. When the response carries an issuer (iss) value, the client must check it against the issuer it recorded, before it sends the authorization code to a token endpoint [S4 §Authorization Response Validation].
4. **The client lists the tools.** From here every HTTP request carries the token in an Authorization: Bearer header, never in the URL's query string [S4 §Token Requirements]. The server must check that the token was issued for it as the audience, and must not accept or pass on any other token [S4 §Token Handling]. The client sends tools/list, and the server returns each tool's name, description and input schema [S3 §Listing Tools]. The list may vary with the authorization on the request, for example showing only the tools the caller's scopes permit [S3 §Capabilities].
5. **The model picks a tool and the client calls it.** Tools are model-controlled: the model can find and call them on its own, based on the context and the user's prompt [S3 §User Interaction Model]. The protocol does not require any particular way of involving the user, but there should always be a person who can deny a call, and applications should show confirmation prompts [S3 §User Interaction Model]. The specification's security principles also ask hosts to get the user's consent before invoking a tool, though MCP cannot enforce them [S1]. The client then sends tools/call with the tool's name and arguments. The server returns the result and sets isError when the tool itself failed [S3 §Calling Tools].
6. **The server asks for more scope when it needs it.** If the token lacks a scope an operation needs, the server should answer HTTP 403 with the error insufficient_scope and list every scope the operation needs in one challenge [S4 §Scope Challenge Handling]. A client acting for a user should then re-authorize for the scopes it already asked for plus the new ones, so it keeps what it had [S4 §Step-Up Authorization Flow]. Clients should limit their retries and stop after a few failed attempts [S4 §Step-Up Authorization Flow].

## Why does Model Context Protocol matter now?

Many MCP servers are reachable with no sign-in at all. In one academic scan, 3,233 of the 7,973 remote servers that answered took no authentication [BI-FIND-2026-0945](https://insights.bostonidentity.com/ref/BI-FIND-2026-0945). Of 119 servers that let outsiders register an OAuth client, 114 accepted any redirect address, which lets an attacker collect an authorization code [BI-FIND-2026-0944](https://insights.bostonidentity.com/ref/BI-FIND-2026-0944). At Black Hat, a research team showed that a malicious server picks the sign-in address a client opens, and turned that into code running as the user [BI-FIND-2026-0746](https://insights.bostonidentity.com/ref/BI-FIND-2026-0746).

The tool descriptions a model reads are a second way in. In the MCPTox study, poisoned descriptions for 353 tools on 45 real servers led agents to misuse tools they already held, at rates up to 72.8% for o1-mini [BI-FIND-2025-0039](https://insights.bostonidentity.com/ref/BI-FIND-2025-0039).

In September, five vendors shipped controls for MCP servers, and three of them route an agent's tool calls through an identity check [BI-FIND-2026-0703](https://insights.bostonidentity.com/ref/BI-FIND-2026-0703). In a February concept paper, NIST's National Cybersecurity Center of Excellence listed MCP among the standards it is considering for agent identity [BI-FIND-2026-0686](https://insights.bostonidentity.com/ref/BI-FIND-2026-0686).

## Where does it fit?

- Compared with OAuth: OAuth is the framework underneath. MCP's authorization uses a chosen subset of OAuth 2.1 and related standards: a protected MCP server is an OAuth 2.1 resource server, and the MCP client is an OAuth 2.1 client [S4 §Roles].
- Compared with Dynamic Client Registration: Dynamic Client Registration lets a client register itself with an authorization server at run time. MCP's current specification deprecates it in favour of Client ID Metadata Documents and keeps it only for authorization servers that lack them [S4 §Overview].
- Compared with Cross App Access: Cross App Access puts the company's identity provider between an agent and each application [BI-FIND-2026-0261](https://insights.bostonidentity.com/ref/BI-FIND-2026-0261). MCP carries it as an opt-in extension, so an employee signs in once at the identity provider instead of approving each server [BI-FIND-2026-0678](https://insights.bostonidentity.com/ref/BI-FIND-2026-0678).
- Compared with OpenPort Protocol: OpenPort is a separate proposal for exposing tools to agents through a server-side gateway, and it requires that an agent never see tools its credential cannot use [BI-FIND-2026-1239](https://insights.bostonidentity.com/ref/BI-FIND-2026-1239). MCP only allows a server to vary its tool list by the caller's authorization [S3 §Capabilities].

## What goes wrong?

- Clients must treat tool annotations, the hints about how a tool behaves, as untrusted unless they come from a trusted server [S3 §Tool]. Tool descriptions can carry attacks too: poisoned descriptions have steered agents into misusing tools they already held [BI-FIND-2025-0039](https://insights.bostonidentity.com/ref/BI-FIND-2025-0039).
- Authorization is optional in the specification [S4], which also says MCP cannot enforce its consent and access principles itself and asks implementers to build them [S1]. In one scan, four in ten remote servers took no authentication [BI-FIND-2026-0945](https://insights.bostonidentity.com/ref/BI-FIND-2026-0945).
- A server must not accept or pass on a token issued for anything else [S4 §Token Handling], so a gateway that hands a person's token to the tools behind it is out of specification [BI-FIND-2026-0191](https://insights.bostonidentity.com/ref/BI-FIND-2026-0191).
- Servers that accept anonymous registration with any redirect address let attackers collect authorization codes [BI-FIND-2026-0944](https://insights.bostonidentity.com/ref/BI-FIND-2026-0944). The current specification has deprecated Dynamic Client Registration, the route those servers use [S4 §Overview].
- Servers must validate all tool inputs, implement proper access controls, rate-limit tool calls and sanitize tool outputs [S3 §Security Considerations]. Clients should show a tool's inputs to the user before calling the server, to avoid malicious or accidental data leaks, and should log tool use for audit [S3 §Security Considerations].
- The specification makes the client a plain OAuth client. Nothing in the token says which person an agent acts for, and delegation is left to optional extensions [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189).

## What should a security leader ask?

1. Which MCP servers can our AI tools reach today, and which of them accept requests with no authentication?
2. Does each MCP server we run check that a token was issued for it, and does any gateway in front of it forward a user's token to other services?
3. Which of our MCP clients still register through Dynamic Client Registration, and when do they move to Client ID Metadata Documents?
4. Which tool calls run without a person approving them, and where are tool inputs and calls logged for audit?
5. When an agent calls a tool for an employee, can our logs show which person it acted for?

## Sources

- [S1] [Specification - Model Context Protocol](https://modelcontextprotocol.io/specification/latest) (Model Context Protocol project, Latest revision)
- [S2] [Versioning and Compatibility - Model Context Protocol](https://modelcontextprotocol.io/specification/latest/basic/lifecycle) (Model Context Protocol project, Revision 2026-07-28 (latest))
- [S3] [Tools - Model Context Protocol](https://modelcontextprotocol.io/specification/latest/server/tools) (Model Context Protocol project, Latest revision)
- [S4] [Authorization - Model Context Protocol](https://modelcontextprotocol.io/specification/latest/basic/authorization) (Model Context Protocol project, Latest revision)

## Evidence

- [BI-FIND-2026-0069](https://insights.bostonidentity.com/ref/BI-FIND-2026-0069): MCP with its community-added OAuth layer is the de facto tool-call protocol, and the panel is unanimous that it is not a policy system: it says how a client gets a token, not who is allowed to do what. The access decision is being pulled up into identity providers and agent gateways that sit above the protocol, which is exactly where Okta, Microsoft Entra and Ping Identity already sit for humans. (from "The rules cover one agent acting for one person; they break when that agent hands work to other agents")
- [BI-FIND-2026-0704](https://insights.bostonidentity.com/ref/BI-FIND-2026-0704): Okta and Ping both shipped client ID metadata documents (CIMD), and CyberArk went the other way. With CIMD, an agent or app names itself with a URL that points to its own metadata, and needs no prior registration. Okta made CIMD a beta for AI agents and GA in preview for OIDC apps on September 16. Ping's Advanced Identity Cloud shipped it on September 29. CyberArk's September MCP feature relies on dynamic client registration (DCR) instead. Okta and Ping also changed their DCR support in the same releases, so neither has dropped it. (from "Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally available")
- [BI-FIND-2026-0945](https://insights.bostonidentity.com/ref/BI-FIND-2026-0945): Four in ten of the remote MCP servers the authors found answered with no authentication: 3,233 of 7,973 servers that responded to an MCP handshake, against 2,428 using OAuth and 2,312 using static tokens or API keys. When the authors sampled the open ones, most were test or demonstration deployments with nothing sensitive in them. One was an internal customer-records server, meant to be private, that let anyone query more than 5,000 customer records (CVE-2025-61510). The count covers hosts that FOFA and Shodan index, so servers those engines miss are not in it. (from "114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sent")
- [BI-FIND-2026-0944](https://insights.bostonidentity.com/ref/BI-FIND-2026-0944): Of the 119 remote MCP servers whose OAuth setups Zhou, Zhang and colleagues could test end to end, all of which register clients dynamically, 114 accepted any redirect address from an anonymous registrant. That lets an attacker register a client with a callback of their own, get the victim to click a genuine sign-in link, and collect the authorization code; seven of the nine CVEs the authors obtained trace to this flaw. Every server tested had at least one flaw, 325 in all. The 119 are what was left of 1,118 servers advertising a registration endpoint after the authors dropped duplicates, unreachable hosts and 316 servers behind enterprise access controls, so the rate describes servers open to outside registration, not MCP as a whole. (from "114 of 119 MCP servers that let outsiders register an OAuth client let them choose where sign-in codes are sent")
- [BI-FIND-2026-0746](https://insights.bostonidentity.com/ref/BI-FIND-2026-0746): MCP's sign-in flow lets the server tell the client where to send the user, and Zhong's team showed that the client passes that whole authorization URL to whatever opens it. A malicious or compromised server turned that into code running as the user: a PowerShell sub-expression through the widely used npm `open` package on Windows, which affected Google's Gemini CLI before version 0.17; file and SMB links that ran local or remote scripts; script injection in browser-based clients; and, in Anthropic's MCP Inspector, a stolen CSRF token that let a request start a local command. The fix in the official TypeScript SDK blocks javascript:, data: and vbscript: links but not file: or SMB, and the specification gives no rule for validating these URLs. (from "An MCP server picks the login URL its client opens, and that was enough to run code on users' machines")
- [BI-FIND-2025-0039](https://insights.bostonidentity.com/ref/BI-FIND-2025-0039): A tool's description is enough to turn an agent against its user. MCPTox wrote poisoned descriptions for 353 tools on 45 real MCP servers and found agents acting on them across the board: o1-mini at 72.8%, DeepSeek-R1 at 70.9% and Phi-4 at 70.2%, down to 8.3% for Mistral, with refusals at 3% or below for every model. The attack needs no extra permission, because the poisoned text tells the agent to use a tool it already holds, such as reading an SSH private key or redirecting email to an attacker's address. The headline rates average three prompt settings, the paper's case counts do not add up, and its claim that more capable models are more vulnerable rests on one model family and breaks within it. No defense was tested. (from "MCPTox: a poisoned tool description steers most agents into misusing the tools they already hold")
- [BI-FIND-2026-0703](https://insights.bostonidentity.com/ref/BI-FIND-2026-0703): Five vendors shipped controls for MCP servers in September. Three of them proxy an agent's tool calls through an identity check, and only Cloudflare's proxy is generally available. Cloudflare's MCP server portals went GA on September 24. Okta's Agent Gateway APIs went to beta on September 16, and Google put MCP behind API Gateway in preview. The other two do different jobs: CyberArk registers MCP servers automatically when they support dynamic registration, and SailPoint lists them through its JDBC and REST connectors. The vendors agree that MCP servers need an identity control point. They do not yet agree on what it should do. (from "Three vendors now route MCP tool calls through an identity check, and only Cloudflare's is generally available")
- [BI-FIND-2026-0686](https://insights.bostonidentity.com/ref/BI-FIND-2026-0686): NIST's National Cybersecurity Center of Excellence framed agent identity in February as a question of applying what exists, not inventing something new. Its concept paper lists OAuth, OpenID Connect, SPIFFE, SCIM, the NGAC access-control standard and MCP as candidates it is "currently considering", and names none of the agent-specific IETF work, such as ID-JAG, the WIMSE drafts or token exchange for agents. It keeps human delegation and logging in scope and leaves agents from outside the enterprise for a later effort. The next steps are a possible project description and a lab-built practice guide, with no date; the comment period closed on 2 April. (from "NIST's agent identity work starts from OAuth, SPIFFE and SCIM and defers agents from outside the enterprise")
- [BI-FIND-2026-0261](https://insights.bostonidentity.com/ref/BI-FIND-2026-0261): Okta's product chief presents Cross App Access as the company's answer for agents: the identity provider sits between an agent and each application, an administrator approves the connection once, and the agent receives a short-lived token with no consent prompt. An agent acting for a person gets the overlap of that person's permissions and its own granted scopes; an autonomous one gets its own identity. The mechanism is the identity assertion grant now in the IETF OAuth working group, and the Model Context Protocol carries it as an extension. Kahn claims more than twenty-five software vendors are adopting it and names only Anthropic. (from "Okta's answer to agent sprawl: put the identity provider between every agent and every app")
- [BI-FIND-2026-0678](https://insights.bostonidentity.com/ref/BI-FIND-2026-0678): MCP now has a stable way for a company's identity provider to stand between an employee's AI client and every MCP server. The client trades the employee's single sign-on ID token at the IdP for an ID-JAG, using OAuth token exchange, and hands that to the MCP server's own authorization server as a JWT assertion; the IdP checks group, role and conditional-access policy before issuing, and the employee sees one corporate login instead of a consent screen per server. The ID-JAG must be addressed to that authorization server, and when the client names a resource the access token must be restricted to it. The extension is opt-in: a server signals support in its authorization-server metadata. (from "MCP's enterprise login extension says IdP revocation is immediate. Its own spec does not back that up")
- [BI-FIND-2026-1239](https://insights.bostonidentity.com/ref/BI-FIND-2026-1239): In OpenPort, tool discovery is part of authorization. The manifest an agent fetches must not show tools or fields its credential is not allowed to use, and error details should be redacted so a denial does not reveal privileged names. The authors argue that tool-exposure protocols such as MCP often leave which tools are visible under which credentials to each server. Delegated OAuth tokens must be mapped to minimal OpenPort scopes and policy windows before any tool access, and the protocol names no identity provider: it "mandates that authorization decisions are made server-side and are auditable." (from "OpenPort holds agent writes as drafts for an operator by default; its own tests check less than its spec asks")
- [BI-FIND-2026-0191](https://insights.bostonidentity.com/ref/BI-FIND-2026-0191): The specification binds every token to one server and forbids passing it on. Clients MUST send a resource parameter naming the server's canonical URI on every authorization and token request, and MUST check the issuer on every authorization response, the two controls against mix-up attacks. Servers MUST validate the audience of every token and 'MUST NOT accept or transit any other tokens'. Tokens travel only in the Authorization header, refresh tokens stay confidential and are never assumed, and offline access should not be advertised. That is the same anti-pattern the six-vendor agent draft names, a tool forwarding the token it received, and it is the rule a token-passthrough gateway breaks: a proxy that hands a person's token to the tools behind it is out of specification, whatever the vendor calls it. (from "MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension")
- [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189): As of the 28 July 2026 revision, an MCP client is an OAuth 2.1 client and an MCP server is an OAuth 2.1 resource server, and that is the whole identity model. Authorization is optional; HTTP transports SHOULD use it, and 'implementations using an STDIO transport SHOULD NOT follow this specification, and instead retrieve credentials from the environment'. The authorization server is out of scope. There is no actor claim, no on-behalf-of flow, no token exchange and no transaction-token profile: nothing in the token says whom the agent is acting for, which is what the Identiverse summit panel meant by 'not a policy system' and what WorkOS meant by the gateway being undefined. Anything of that kind is routed to the new extension model, optional and additive, in a separate repository. This page is the baseline against which the summit's forecast, that the specification gains a delegated-identity mechanism within a year, will be graded. (from "MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/mcp. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
