The Model Context Protocol (MCP) is an open protocol that lets an AI tool call an outside system. For governance work, that means people and AI agents can ask a governance platform questions from an AI assistant or a code editor without opening the platform. Atlan’s context layer for AI exposes owners, certification status, glossary terms, lineage and applicable policies through a hosted MCP server. Any tool that changes something returns a preview and waits for your approval. The server passes your question or proposed change to the platform, and the platform supplies the answer.
“MCP data governance” gets used two ways. One is governing which MCP servers agents may use, which an MCP gateway and an MCP registry address. The other is doing governance work through MCP, and that is the subject here.
Gartner published its 2026 Critical Capabilities and Magic Quadrant for Data and Analytics Governance Platforms on 12 November 2026. To see what each report covers, read our Critical Capabilities page and Magic Quadrant explainer.
| Quick facts | |
|---|---|
| What it is | Governance questions and proposals carried by an open protocol, from the tools people already use |
| Best for | Stewards, data owners, AI leads and builders who need governed context mid-task |
| What an agent can read | Certification status, owners, glossary terms, lineage and the policies that apply to an asset |
| What needs approval | Any metadata change on Atlan’s server: it returns a preview and waits for your approval |
| Spec revision checked | 2026-07-28 (as of November 2026) |
| First check | Whose identity the call runs as, since that sets what the answer can reveal |
What is MCP, and why does governance work move into the tools people already use?
MCP is an open protocol that lets an AI tool call an outside system. That makes it possible to ask governance questions in the chat or editor where the work is happening.
What an MCP server contains
A server exposes tools that the model can call on its own, and resources that the host application chooses to include as context. The MCP specification calls tools “model-controlled,” while resources are application-driven. That split matters for governance, because the tool is the part a model can choose to call during a task, so what it returns and what it may change need controls.
Example: a steward’s question in chat
Governance questions rarely arrive on schedule. An analyst building a dashboard wants to know who owns a table and whether anyone certified it, and an agent drafting a query needs to know which policies apply before it touches a column. Reading a policy helps the agent plan, but it does not enforce anything. When you sign in with OAuth (the standard sign-in protocol), each call can reach only what your own account is allowed to see.
How MCP differs from a separate plugin per tool
MCP adds one path that an assistant or agent can call from any client that supports the server’s transport and sign-in method, so the question does not need a separate plugin per tool. For the wider case, see why MCP matters for AI agents and how MCP architecture fits together.
Why the answer depends on the context layer
The answer is only as good as what sits behind the server. Atlan’s context layer for AI holds the governed definitions, owners and certification that the server reads, which is why context engineering and a trustworthy MCP connection to business context belong together. A server returns only what the platform holds, so answers are only as reliable as the metadata behind them.
What governance tasks can Atlan’s MCP server serve?
Atlan’s MCP server returns context an agent can read, such as definitions, owners, certification status, lineage and policies, and lets it propose changes.
Which questions the server answers
Atlan’s docs list the fields and tools behind each question below. Other MCP servers follow the same pattern, but the field names here are Atlan’s. The table lists what the agent reads, what the platform has to hold for the answer to be right, and where each task stops.
| Question | What the agent reads or proposes | What the platform must hold | Where it stops |
|---|---|---|---|
| Who owns this asset? | Owner users, owner groups and admin users on the asset | Named owners, assigned by people | Reports the owners; who owns an asset is set in the platform |
| Is it certified? | Certificate status and its message | A certification someone actually set | Reports the status; a proposed change waits for your approval |
| What does this term mean? | Glossary terms attached to the asset | Governed definitions in a glossary | Returns the term; new terms follow the platform’s approval workflows |
| Where does it come from, and who is downstream? | Lineage traversal in both directions | Lineage captured from the data stack | Shows the path; people judge the impact |
| Which policies apply? | The policies, from personas and purposes, that apply to the asset | Personas and purposes defined by admins | Reads which policies apply; reading is not enforcing, and the platform holds the policy |
| Can this metadata change? | A proposed change with a preview | Permission to update the asset | Waits for your approval before writing |
Use cases and references
Each row maps to attributes listed in the Atlan MCP asset attributes reference and the tools reference. The governance and quality use cases add workflows such as recertifying assets, propagating PII tags to parent assets, adding ownership context to quality alerts and scoring assets for AI readiness. Those are reads, plus proposals that wait for your approval. Lineage goes deeper in MCP for data lineage and root cause analysis with MCP.
What the platform must hold
The right-hand column carries the point. An agent can report that a table is certified only if someone set that status, and the same holds for data certification and for owners and stewards. Agents also need rules written so machines can read them. The catalog side of the same story, including why governed metadata decides answer quality, is covered in MCP-connected data catalog.
What should you check before trusting an MCP server with governance work?
Before an agent touches governance metadata, check four things on the server: identity, scopes, approval and audit. Then check the specification revision it speaks.
Identity
The MCP specification sets out what to expect, and Atlan’s security documentation shows how one server meets each point. Confirm each answer in a trial. Identity comes first. The MCP security best practices say a server must not accept tokens that were not issued for it, and they list token passthrough as an anti-pattern that makes incident investigation and auditing harder. For people, per-user OAuth works better because the token carries that person’s own identity. On Atlan’s server each person signs in with their own account, so every call runs as the signed-in user, while a shared API key acts as one shared identity and suits automation better than people. Agent identity is the topic to read next.
Scopes
The same security best practices recommend a progressive, least-privilege scope model: minimal discovery and read scopes first, elevation when a privileged operation is attempted, no wildcard scopes, and elevation events logged. Atlan’s documentation states that users can only search, view or update metadata they are already authorized to access, and that MCP never widens what a login permits. Tool exposure is allowlisted per tenant and fails closed, per the same security reference. Related reading: AI agent access control and zero trust for data governance.
Approval
The MCP specification on tools says a person should always be able to deny a tool invocation, and it treats tool annotations as untrusted unless they come from a trusted server. Atlan’s write tools return a preview and wait for your approval, as the tools reference describes, so the person who made the request confirms the change. If your policy needs a steward’s sign-off on agent changes, ask your admin how to set that up before rollout. The remote server exposes both read and write tools by default, and a read-only mode, requested through Atlan support, removes every write, admin and lifecycle tool while keeping search and lineage. Starting read-only is the safer first setting, and enabling writes later should be a decision someone makes on purpose.
Audit
The specification says clients should log tool usage for audit, so ask what the server records as well: the identity behind each call, failed and denied calls, and how long the records are kept and read. Atlan logs every tool call as a structured event with sensitive arguments and responses redacted, keeps those logs on its central observability pipeline’s retention rather than a separate MCP setting, and shows metadata changes in the asset’s activity history, as covered in its MCP security reference. Confirm that this record meets your own audit needs.
Also look at what the answers themselves expose: descriptions and glossary text are written by whoever last edited them, so treat them as input rather than instructions, and returned metadata goes to whichever model your client uses.
Specification revision
Ask which specification revision the server speaks. The current revision is 2026-07-28, and some servers and clients still implement earlier revisions (2025-11-25 and before), so check which one yours speaks.
Setups to avoid
Four setups to avoid: a shared service key behind a user-facing assistant, wildcard scopes, write tools left on without anyone deciding (Atlan’s remote server ships with them on, so ask for read-only mode first), and a third-party server’s tool descriptions trusted without review. Portal-level controls such as access policies, per-tool scoping and audit logs help with the first meaning of MCP governance, as a gateway provider’s MCP governance documentation describes, and MCP gateway vs context layer explains where each earns its place. A written guardrails checklist and role-based access in a context layer cover the rest.
What does Gartner’s 2026 research say about MCP?
A key finding in Gartner’s Critical Capabilities for Data and Analytics Governance Platforms (12 November 2026) mentions MCP. It says platforms are being used “through Model Context Protocol (MCP) server experiences as a way to interact with the platforms in a frictionless way, using regular workflows.” You can download the free 2026 Gartner Magic Quadrant for D&A Governance Platforms from Atlan, and Atlan’s own pages cover the Critical Capabilities report and the Magic Quadrant.
The finding is about reaching platforms through regular workflows. It says nothing on identity, approval or audit, so the four checks above still apply.
How does Atlan approach MCP for governance work?
Atlan hosts an MCP server so AI tools can read enterprise context from its context layer for AI.
What the hosted server offers
From the client you can search, trace lineage, read governed definitions and glossaries, curate metadata and check data quality. Atlan hosts the server for every tenant, so there is no server to install or run, and admins decide who can use it in Labs, the settings area where Atlan manages its AI features. The tools reference lists 41 tools in nine categories, some gated per tenant.
Where control sits
Four controls apply. OAuth keeps each person’s own permissions. Atlan lets each tenant choose which tools are available. Every tool call is logged as a structured event. Changes return a preview and wait for your approval.
Related reading
The context layer for data governance teams gives stewards one place to define what agents read, the context layer reference architecture shows where the server fits, and agent context layer design covers how to shape what an agent receives. What is Atlan MCP covers the product itself.
How can you try MCP with a governance platform?
Trying it takes four steps: connect a client, sign in as yourself, ask read-only questions first, then test one proposed change and approve it.
The four steps
- Enable the server; on Atlan, an admin does this in Labs. The setup overview lists the endpoint and the clients it supports.
- Connect a client with OAuth where the client allows it. Some clients support only OAuth and others only an API key, so check the matrix before you pick one. Connecting an AI assistant to your context layer walks through one path.
- Ask three read-only questions about an asset you know well: who owns it, whether it is certified, and which policies apply. Compare the answers with what the platform shows. Then sign in as a second user with fewer permissions and ask about an asset that user cannot see, and confirm the server declines without describing it.
- Propose one change, read the preview, approve it, then open the asset’s activity history to see the change recorded.
Before a wider rollout
Connecting an external assistant also brings that client’s own data handling into scope, so review it. Ask the admin about read-only mode before a wider rollout, and see the MCP server implementation guide and how to build MCP servers for enterprise data if you are building one of your own. For programmatic jobs, see MCP vs API.
Where MCP stops and the platform’s workflows start
In Atlan’s design, an MCP server carries questions and proposals, and the owners, approvals, tasks and policies stay in the governance platform.
What stays in workflows
Atlan’s stewardship workflows route data access requests, metadata changes and new glossary and tag entries through approvals, and those approvals stay in Atlan’s workflows and tasks. MCP handles search, reading owners, certification, terms, lineage and applicable policies, and proposing changes with a preview.
Stewardship behind the answers
Where an owner or certification field is empty, the agent has nothing to report, so stewardship work sets the quality of every answer. These pages cover how to do it: a data governance operating model, embedded governance and lessons from failed implementations cover how. Access requests in a ticketing flow have their own pattern in running data access requests in a ticketing flow.
Get the free 2026 Gartner Magic Quadrant for Data and Analytics Governance Platforms
Get the Magic QuadrantFAQs about MCP for data governance
Readers most often ask ten questions about MCP and governance work.
1. What is MCP governance?
The phrase has two meanings. One is governing which MCP servers and tools agents may use, which gateways and registries address. The other is doing governance work through MCP, such as asking for an owner, a definition or a certification status.
2. What is an MCP for data?
An MCP server that exposes data tools and metadata to AI clients. The client calls the tools to search, read lineage or look up definitions, and the server should apply its access controls to each call. Atlan’s hosted MCP server is one example.
3. Can an agent change governance metadata through MCP?
Yes, on Atlan’s server, as a proposal. Tools that change anything return a preview of the change and wait for your approval before writing. The person who asked approves the preview, and the change then appears in the asset’s activity history.
4. Does MCP respect my existing permissions?
With OAuth, yes. Each person signs in with their own Atlan account, so tools run with that person’s permissions, and Atlan’s documentation says MCP never widens what a login permits. A shared API key acts as one shared identity, so reserve it for automation and service accounts.
5. Does MCP replace governance workflows and approvals?
No. On Atlan, access requests, steward tasks, approvals and policy definitions run in the platform’s workflows. MCP is a front door for questions and proposals into the same platform.
6. What should I log when agents use MCP?
Log every tool call, the identity it ran as, and any step up to broader permissions. The MCP specification says clients should log tool usage for audit. Atlan logs each call as a structured event and records metadata changes in the asset’s activity history.
7. How do I control which data categories agents can access?
Through the permissions of the person or service account the agent runs as. Atlan’s MCP server passes that identity to Atlan, which applies the same persona, role and domain policies as the Atlan UI. Tool exposure is also set per tenant, and read-only mode removes every write tool.
8. Can I enforce different data governance policies for different teams?
Atlan applies persona, role and domain policies to each MCP call, so teams whose people hold different personas or domains see different results through the same server. Define those policies in Atlan, not in the MCP client. Each service-account token must have exactly one persona.
9. Does data governance for AI require a separate tool from our existing DLP?
An MCP server does not replace data loss prevention. It returns metadata and applies the signed-in person’s permissions to each call, so it governs what an agent can read in the platform, not traffic leaving your network or other systems. Keep your existing DLP for those, and treat the two as layers.
10. What happens if a filter or policy misses something?
Atlan’s documentation says MCP never widens what a login already permits, and that if the server cannot resolve a user’s workspace role it narrows scope rather than exposing more. A missed policy in Atlan itself is still a gap in the platform, so test with a lower-permission user and review the audit log of tool calls.
Sources
- Tools, Model Context Protocol specification, revision 2026-07-28
- Resources, Model Context Protocol specification, revision 2026-07-28
- Security best practices, Model Context Protocol, revision 2026-07-28
- MCP governance, a gateway provider’s Agents documentation
- Atlan MCP setup overview, Atlan documentation
- Atlan MCP tools, Atlan documentation
- Atlan MCP security, Atlan documentation
- Atlan MCP asset attributes, Atlan documentation
- Atlan MCP governance and quality use cases, Atlan documentation
- Automate data governance, stewardship, Atlan documentation
- Gartner, Critical Capabilities for Data and Analytics Governance Platforms, 12 November 2026, Sarah Turkaly Et Al.
Gartner does not endorse any company, vendor, product or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner’s business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.
Gartner and Magic Quadrant are a trademark of Gartner, Inc., and/or its affiliates.