
MCP Governance: a practical guide to servers, access and approvals
MCP Governance in practice: decide which servers and tools to trust, who can use them and which actions need review, then test and keep evidence as you grow.
The decision in brief
MCP Governance comes down to three decisions, each with an owner, a record and a way to undo it: which servers and tools to trust, who can use them with which data, and which agent actions need review. Start with an inventory and one connection, test allowed and refused requests, then exercise review and recovery. Expand only when the team can explain both the permission decision and the execution result.
What MCP Governance covers
MCP Governance is the set of decisions and controls for how AI agents use MCP servers at work. It comes down to three decisions: which servers and tools you trust, who can use them with which data, and which agent actions need a person’s review before they run.
Each decision needs an owner, a record and a way to undo it. The detailed guides cover each one: server governance for approving and reviewing servers, access governance for people and data, and approval workflows for individual actions.
| Decision | Usual owner | Record | How it is undone |
|---|---|---|---|
| Which servers and tools to trust | Platform or IT team | Inventory entry and reviewed tool list | Disable the connection and revoke its credentials |
| Who can use them, with which data | Connection owner with security | Team assignments and access rules | Remove membership or narrow the rules |
| Which actions need review | Business owner with security | Policy version and review decisions | Publish a stricter policy; approvals expire |
Define the scope of governance
MCP makes capabilities available through a common protocol. Governance determines who may use them, which changes need review, and what evidence operators need afterward. Those controls can span the client, gateway, server, and upstream provider.
A gateway gives teams one place to apply shared controls, but only for requests it receives. Inventory direct connections and alternative credentials too. Do not mistake a successful gateway test for control over every way a user or agent can reach the same data.
Inventory access and assign owners
Begin with one concrete use case. Record the client, requester, connection, upstream account, environment, tools, and data involved. Name the person responsible for granting access, reviewing changes, and resolving failures.
- Clients and requesters: identify people, automation identities, and how access is revoked.
- Connections: record the upstream account, credential owner, scopes, and environment.
- Tools: distinguish reads from writes and identify sensitive data, bulk operations, and external effects.
- Alternative paths: list direct provider credentials and connections that bypass the gateway.
- Operations: assign reviewers, incident owners, and a process for reconciling uncertain writes.
Configure policy and review tools separately
In Durin, requests must pass identity, membership, team, connection, credential, reviewed-tool, resource, and inspection checks. Default allow does not override a failed check. Administrators can deny all tool access and independently allow reviewed writes, require another administrator, or deny writes.
Team policy can narrow organization policy, and the same controls apply in sandbox and production. Custom rules can target connections, tools, resources, environments, people or teams; when rules match, deny takes precedence over approval, then allow. There is no data-classification condition, so map your operating rules to the conditions that are actually available.
Unreviewed tools remain blocked until tool review is complete. An exact write approval cannot substitute for that review, make an unavailable connection ready, or grant missing provider permissions.
Test one connection before expanding
Start with the Durin sandbox or another reviewed connection. Begin with a low-sensitivity read and inspect the requester, decision, and result. Test a refused request as well, such as a resource outside the allowed scope.
For the write exercise, configure the approval requirement and involve a different active administrator. Confirm that the initial request and the approval itself do not execute the action. Retry the exact call before expiry, then check the execution result.
Test revocation, changed arguments, expired grants, and tool-definition changes. Use a controlled test to exercise an ambiguous write result and confirm that it is not automatically repeated. Expand the rollout using the same checks.
Keep useful evidence with a clear retention policy
Record enough context to explain the requester, connection, tool, policy decision, approval where required, and execution outcome. Authorization to run and confirmed completion are different facts. Assign someone to reconcile uncertain writes with the upstream system.
Decision records do not require retaining every prompt or response. Durin defaults conversation capture to metadata only. When an administrator enables redacted capture, new captures are retained for 7, 90, or 180 days, or 360 days on Enterprise; workspaces with an older 1–30-day setting keep it until an administrator selects a supported period. These retention periods are separate from audit, account, and billing retention.
Use the trust materials to review control scope and supporting evidence. Certifications, independent assessments, and contractual commitments require their own evidence.
Use these questions at each rollout review
Revisit the checklist when a client, connection, credential, tool, or policy changes. Governance is an operating process as well as a gateway configuration.
- Can we identify the requester and the owner of every enabled connection?
- Have we tested both permitted and denied access to the intended resources?
- Are tool review and per-action approval treated as separate decisions?
- Does the configured write policy match the team’s requirements?
- Have we tested expiry, exact retry, revocation, and uncertain-outcome handling?
- Do we know which access paths bypass the gateway?
- Are evidence access, retention, and incident ownership documented?