Implementation Guide | Healthcare

    Controlling Which MCP Tools an n8n AI Agent Can Call When Workflows Carry PHI

    Control MCP tool access in two layers. At design time, expose only the tools each workflow needs, give MCP servers least-privilege credentials, and keep PHI and non-PHI paths in separate workflows. At runtime, enforce policy on every tool call (allow, redact, block, rate limit, escalate for human approval) and record each decision. The design-time layer limits what the agent can see. Only the runtime layer governs what it does with PHI, and only recorded activity satisfies 164.312(b).

    1. 1

      Tool list on the MCP client node

      Constrains what the agent can discover and call. Set at design time.

    2. 2

      MCP server credentials

      Constrain what a tool can do once called. Set in the server and the credential store.

    3. 3

      Workflow structure

      Constrains which agents ever see PHI.

    4. 4

      Network egress

      Constrains where tool traffic can go.

    5. 5

      Runtime policy on each call

      Decides per call whether to allow, redact, block, rate limit or escalate.

    6. 6

      Evidence store

      Records what was permitted and what happened.

    7. 7

      Allow

      Permit calls to approved tools, with approved destinations, from approved workflows. Record the decision.

    8. 8

      Redact

      Remove or mask PHI in arguments or responses when the tool or the model does not need the identifiers.

    9. 9

      Block

      Stop calls to unlisted tools, to unapproved URLs or domains, or with arguments outside the expected shape.

    10. 10

      Rate limit

      Cap call volume per tool or per workflow to limit bulk retrieval or repeated attempts.

    11. 11

      Escalate for human approval

      Hold consequential actions, such as writing to a record or sending data outside the organization, until a named person approves.

    Evidence an auditor will ask for

    • The approved tool list and policy version in force for each workflow, with change history.
    • The service identity and credential scope behind each MCP server.
    • Per call: timestamp, workflow, agent, tool, destination and the policy decision taken.
    • Whether PHI was redacted, and which rule applied.
    • Blocked and rate-limited calls, which show the controls operating.
    • Human approval records: who approved, when, and what was approved.

    Apply the design-time restrictions first

    Design-time controls are cheap, deterministic and worth doing even if you add runtime enforcement. They shrink the set of decisions the runtime layer has to make.

    Start with scoped tool selection. Configure each MCP client node to expose only the tools that workflow needs, rather than everything a server advertises. Treat the exposed list as a reviewed artifact: a new tool added to a server should not reach an agent without a change review.

    Next, apply least privilege to credentials. Give each MCP server a dedicated service identity with only the permissions its tools require, and avoid one broad credential shared across workflows. Review who can use each credential in the workflow platform and which tools can reference it. Include these service identities in your access authorization and modification procedures under 164.308(a)(4).

    Finally, separate PHI and non-PHI paths. A workflow that handles PHI should be a distinct workflow with its own narrow tool list, not a branch of a general-purpose agent. Keep tools that reach external destinations, such as web requests or messaging, out of PHI workflows unless a specific approved use requires them.

    • Expose the minimum tool list per workflow and review changes to it.
    • Use one scoped credential per MCP server or purpose.
    • Restrict which users and workflows can use each shared credential.
    • Place PHI handling in dedicated workflows with their own tool lists.
    • Limit network egress from the environment running the MCP servers.

    Where design-time controls stop

    A configured tool list tells you what the agent was allowed to see. It does not evaluate the arguments of a particular call, redact PHI from a request or a response, or produce a record of what the agent did. Those are runtime properties, and 164.312(b) is about runtime: it calls for mechanisms that record and examine activity, not for a configuration that implies it.

    Configured restrictions can also fail in the tool layer. A published vulnerability, CVE-2026-59207, describes an AI Agents feature that did not enforce the Allowed HTTP Request Domains restriction set on a credential when an MCP tool was pointed at an arbitrary URL. A member-level user with use-only access to a shared credential could send its secret to an external server they control. The entry lists fixed versions 2.27.4 and 2.28.1. Confirm the affected product in the full record and check your deployed version against it.

    The lesson is general. A domain allowlist on a credential is only as strong as the code path that checks it. Pair it with egress limits and with a check at the point where the call is made, so that one bypassed setting does not leave PHI or secrets with no second control.

    Enforcing policy in the execution path

    You can build per-call policy and evidence yourself with a proxy in front of your MCP servers, or you can use a product built for it. Trussed is one option. It enforces policy in the execution path for tool calls, including calls to MCP servers, with the actions allow, redact, block, rate limit and escalate for human approval. Enforcement happens at runtime, not through review after the fact, and it generates audit-ready evidence mapped to HIPAA. Trussed can be deployed self-managed in your own cloud, with configurable retention and data residency, and runs in production with healthcare and pharma organizations.

    Its scope is individual interactions and tool calls. It does not replace design-time scoping, least-privilege credentials or workflow separation, which remain your responsibility. Whichever approach you choose, test it against a simple question: for any tool call last month, can you show the policy that applied and what the decision was?

    A proposed amendment to the HIPAA Security Rule was published in the Federal Register in January 2025. As of November 2026, it should not be treated as current law, and the controls above are framed against the existing 164.312 text.

    Four layers of tool access control

    Tool selection

    Limit which MCP tools the agent node exposes.

    Credentials

    Scope MCP server and workflow credentials to least privilege.

    Workflow separation

    Keep PHI and non-PHI paths in different workflows.

    Runtime policy and evidence

    Decide and record each tool call.

    See how per-call enforcement and evidence work for MCP tools

    Walk through policy on individual tool calls and the audit-ready evidence it produces for PHI workflows.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.