What you’ll take away
- Approve a specific server build, transport, client configuration, tool set, and business use, with the deployed protocol revision recorded.
- Trace authenticated caller permissions separately from the server’s upstream credential; a confirmation prompt does not enforce backend authorization.
- Require evidence for permitted and denied operations, assign fixes to the responsible component, and revisit approval when the connector changes.
In this guide
Approve a business connection with a defined purpose
A fictional regional distributor wants its purchasing assistant to look up approved suppliers and prepare draft orders. The team proposes an HTTP MCP connector to the purchasing system. The pilot should allow a requester to create drafts for their assigned branch; submitting or releasing orders remains outside the approved use. A successful demonstration shows that the connector can perform useful work. The approval review must establish whether its access and operating limits match that intended use.
Model Context Protocol, or MCP, standardizes how AI applications exchange information and use capabilities exposed by servers. Its architecture distinguishes the host application, the MCP client within that application, and the server providing capabilities. MCP does not dictate how the application uses its model or manages the resulting context. [1] For this review, identify those components and the purchasing API behind the server. Also identify which model services may receive returned supplier or order information.
Create an approval record for the actual deployment: publisher, server build, host and client versions, enabled tools, transport, data scope, and operating owners. This guide references the published 2026-07-28 protocol revision. Your deployed software may support another revision or additional extensions, so record what is implemented and test against the applicable requirements. A product name or registry entry is not enough to establish what code will run or what permissions the proposed configuration will receive.
One request crosses several boundaries.
Look up a supplier, then prepare a branch-specific draft order.
- Host + MCP client
Carries the requester’s identity and presents the proposed draft.
- HTTP MCP server
Validates the caller, then checks the tool, branch, and arguments.
- Purchasing API
Enforces the upstream role’s permitted operations.
Credential intended for this MCP server.
Separate upstream credential; limited authority.
Choose checks that fit the transport
The authorization specification makes authorization optional for MCP implementations. Where supported, HTTP implementations should follow its authorization framework; stdio implementations should instead obtain credentials from their environment rather than use that flow. [2] Those protocol choices do not decide whether anonymous access is appropriate for your business data. The distributor’s proposed HTTP deployment requires authenticated, authorized workforce access as a business requirement that the review must verify.
Local deployment and stdio are not interchangeable descriptions. A server on a laptop can expose an HTTP endpoint. The current Streamable HTTP specification requires Origin validation and recommends localhost-only binding for local servers and proper authentication for connections. [3] Review the endpoint that actually exists, including its binding, reachable networks, and intended callers. Do not assume that a local address automatically protects it from every application or browser running on the device.
For a stdio deployment, focus on the process the host launches, its installation source, inherited credentials, file access, and outbound connections. For HTTP, include the network endpoint and applicable authorization flow. If a gateway or proxy connects the two, review both sides and the authority to start a process. The approval record should state which deployment was examined. Moving the same server from a developer’s machine to a shared service changes the boundary and needs a fresh review of the affected controls.
Trace the caller and the upstream credential separately
For protected HTTP servers using MCP authorization, the specification requires validating access tokens and rejecting tokens issued for another resource. [2] The MCP security guidance separately forbids token passthrough: the server must not forward the incoming MCP access token unchanged to the purchasing API. [4] These checks establish a credential boundary, but do not establish permission to read a particular supplier record or create an order for a particular branch. The application still needs an enforced decision for the requested resource and action.
In the fictional distributor, the server calls the purchasing API with a separate server-held credential. Determine exactly what that credential can do and whether the upstream system sees the individual requester or only the service account. The proposed design needs a reliable mapping from the authenticated caller to their branch permissions. A client name, tool argument, or model-generated statement about someone’s role is not evidence of their entitlement.
Ask the server and purchasing-system owners to identify the enforcement points. The server must evaluate the request against the caller’s permitted branch, and the upstream credential should be restricted to the required draft operations. If the backend can enforce additional per-user restrictions, include those in the design. Record any authority the shared credential retains beyond the approved tools. An assistant that displays only two tools can still depend on a credential with much broader consequences if the connector is misused.
Write a reviewable record for each tool
Start with the actual tool inventory, including names, descriptions, accepted arguments, returned fields, and side effects. For each tool, record the authenticated caller, upstream credential, allowed resources, approval requirement, and recovery options. The fictional lookup_supplier tool reads approved supplier details. Its record should identify which fields are returned and whether access is branch-specific. Calling a tool read-only describes its effect on the source system; it does not resolve whether the returned information is appropriate to disclose.
A proposed record for create_draft_order is more specific: a purchasing requester may create a draft for their assigned branch using an approved supplier and validated line items. The server uses a draft-only purchasing role. The host shows the exact supplier, branch, items, quantities, and total for human confirmation before the call. The server checks the caller’s branch and permitted operation independently. These are proposed requirements for the fictional pilot, not claims about controls already implemented.
Attach acceptance cases to that record. An allowed request should create the intended draft and return its identifier. A different branch, an unauthorized supplier, or a request to submit or release an order should be rejected through the relevant enforcement point. Check the resulting purchasing state, not just the assistant’s wording. Define how duplicate requests and cancellation affect drafts, and who can clean up test records. This gives engineering a concrete contract to implement and reviewers a clear basis for approval.
Make the permitted action testable.
- Tool
- create_draft_order
- Requester scope
- Only the requester’s assigned branch
- Upstream identity
- Separate server-held credential with a draft-only role
- Human confirmation
- Exact supplier, branch, items, quantities, and total
- Allowed test
- Create an agreed own-branch test draft
- Denied tests
- Other-branch draft and order release attempts
lookup_supplier should return only approved supplier details.
Review what descriptions and results can influence
The tools specification says clients must treat tool annotations as untrusted unless they come from trusted servers. It also recommends human involvement with the ability to deny tool calls. [5] A hint that a tool is harmless is therefore not independent evidence of its behavior. Review the implementation and observed effects appropriate to the risk. A confirmation screen supports user oversight; it does not replace checks at the server or purchasing API.
Include controlled tests of misleading instructions in tool descriptions and returned content. For the distributor, a synthetic supplier note could attempt to redirect the assistant toward an unapproved action. The review should observe how the chosen host handles that content and whether the server still enforces the allowed operation. Keep the exercise inside an authorized test environment with synthetic records. The purpose is to identify which component contains a failure and whether the business restriction survives it.
Follow returned information into the host as well. Determine what becomes model context, what appears in the interface, and what is retained in logs or conversation history. Minimize fields to what the purchasing workflow needs. Review each supported host configuration because user confirmations and context handling can differ. If tool definitions or behavior change after approval, refresh the inventory and repeat the affected checks before assuming the earlier review still applies.
Confirm the draft. Enforce the limits.
Show the actual target and contents before confirmation.
Check requester branch and the draft-only upstream role.
Keep the approval tied to the action executed. Retain the resulting draft and trace.
Check installation, secrets, and outbound reach
The MCP security guidance addresses local server compromise and requests sent to unintended destinations, including through authorization discovery. [4] That makes the installation path and outbound network behavior part of the review. Establish which process follows a URL or launches a command; it may be the client, a proxy, or the server. Assign the check to that component rather than treating every issue as a defect in the MCP server itself.
For an installed server, review the publisher, selected package or image version, dependency handling, launch command, and update mechanism. Determine which identity runs it and what files, secrets, and network destinations that identity can access. For a hosted connector, examine the deployed artifact and secret-delivery arrangement. Prefer a controlled release process with a named owner over an installation instruction that silently fetches different code later. A reviewed package still needs appropriate runtime restrictions.
For the purchasing connector, document permitted upstream destinations and how unexpected destinations are blocked. Test the agreed boundary using controlled endpoints rather than probing unrelated internal systems. Check that diagnostic output and error handling do not expose credentials, and that logs retain enough context to investigate a tool call without unnecessarily copying sensitive records. Assign credential rotation, access removal, and connector shutdown to operators who can actually perform them. These operating actions need verification as well as a written procedure.
Ask for a test record that explains the outcome
Prepare a test deployment with representative synthetic data and accounts in different roles or branches. Agree on authorized targets, permitted actions, cleanup, and stop conditions before execution. A third-party purchasing service may impose its own testing conditions, so confirm the allowed environment and methods. The test plan should identify the server build, client configuration, credential model, and deployed revision. Otherwise a later reviewer cannot tell whether the evidence describes the connector awaiting approval.
For each case, retain the expected result, observed result, affected component, and relevant trace. In the fictional draft-order case, evidence should connect the authenticated requester, validated branch, selected tool, upstream operation, and resulting draft or denial. If a different branch’s draft is created, classify that as a failed boundary and identify where authority was lost. An assistant saying that it refused the request is insufficient if the backend action already occurred.
Keep protocol requirements and business acceptance criteria separately identifiable. Mark tested requirements with their result and explain why an item is not applicable. Use a distinct status for untested or blocked work; absence of evidence is not a pass. Record the limits of any sampling and which clients or upstream integrations were excluded. After a fix, replay the affected denial case and an allowed business case so the review checks both the restriction and the intended workflow.
Make approval specific and give it review triggers
Present the decision as an operating choice with conditions. The distributor might approve supplier lookup for a limited group while holding draft creation until branch enforcement is verified. Alternatively, a pilot could remain limited to synthetic data while a credential issue is corrected. These are possible decisions, not reported outcomes. The responsible business and security owners should understand what remains unknown and which conditions must be met before expanding access.
Record the approved build, deployment, hosts, users, tools, data scope, and upstream privileges. Link open issues to an owner, disposition, and review date. A temporary exception must remain visible in the approval record; it does not change a failed test into a successful one. Also define how to disable the connector and who can make that decision. If a problem emerges, the team needs a usable response path rather than an approval document nobody can act on.
Revisit approval after a protocol or SDK change, a new tool, broader credentials, another host application, a new upstream service, or a material configuration change. Assess which evidence remains valid and retest the affected boundaries. Protocol conformance supports interoperability and security requirements within the assessed scope; it does not establish that every business action is appropriate or that the deployment is free of vulnerabilities. Keep that distinction clear when reporting readiness to leadership.
Scope the assessment around the decision you need
Bring the candidate server, supported clients, deployment configuration, tool inventory, sign-in arrangement, and upstream permission model to the scoping discussion. Provide source where available and explain what access a reviewer can obtain when it is not. A test environment and accounts representing different permissions make it possible to examine meaningful boundaries. Identify the specific release decision: approving a third-party installation, allowing an internal pilot to handle real information, or expanding an existing connector’s authority.
Source Point Security’s MCP Server Security Assessment reviews applicable protocol requirements, sign-in and credential handling, tool misuse, upstream authority, and local or network exposure. Deliverables include a requirements checklist, a tool-by-tool table, findings with reproduction evidence, and fixes assigned to the server, authorization service, assistant, or gateway. The engagement can include scoped retesting after changes. It reports the controls tested and the assessment limits; it is not protocol certification or a guarantee of safe operation.
If the question extends beyond one connector into tool chains and the wider agent workflow, SPS’s Agentic AI Security Assessment can address that broader scope. Keep the two decisions connected: a well-constrained connector still operates inside an application with its own permissions and data handling. The useful outcome is a clear approval record with evidence, unresolved questions, responsible owners, and concrete conditions for the next expansion of access.
Review the connector before expanding its access
Bring your MCP server, client configuration, tools, and upstream permission model. We can scope an assessment around the evidence your approval decision needs.
Discuss MCP Server Security AssessmentReferences & further reading
- Model Context Protocol, Architecture overview (2026-07-28)
- Model Context Protocol, Authorization specification (2026-07-28)
- Model Context Protocol, Streamable HTTP specification (2026-07-28)
- Model Context Protocol, Security Best Practices (2026-07-28)
- Model Context Protocol, Tools specification (2026-07-28)
