An agent can start a task with a safe-looking tool menu and finish it after that menu has changed. A connector may add a new write operation, a familiar tool may gain an optional parameter that broadens its reach, or its description may begin steering the model toward a different workflow. If the application treats “this is the same server” as enough identity, the agent can use a capability no reviewer inspected when the task began.
The practical move is to create a tool-surface fingerprint whenever an agent connects to a server or loads a specialist tool group. Record the full discovered catalog in a canonical form, calculate a digest, and bind that digest to the session, approval, or proposal. Before a consequential call, ensure the current catalog still matches the reviewed fingerprint. A material change is not a background refresh; it is a stop condition that requires a fresh inventory and risk decision.
MCP defines tool discovery through tools/list, including each tool’s name, description, input schema, and optional output schema.[1] It also allows a server to announce that the list has changed via notifications/tools/list_changed.[1] Those are useful protocol features, but not an authorization system. MCP specifically says clients must treat tool annotations as untrusted unless they come from trusted servers.[1] OpenAI likewise warns that a remote MCP server can exfiltrate sensitive context and says developers must trust every remote server they use; its integration can require explicit approval for tool calls.[2] The operational implication is to review the actual callable surface, not just a server label or a readOnlyHint.
Make the catalog an input to approval
A server URL identifies where you connected, not the exact operations the model can invoke today. Review the canonical inventory: server identity, tool names, descriptions, schemas, and declared annotations.
Canonicalize and hash it outside the model: sort tools by exact name and object keys, preserve meaningful array order, and use a fixed encoding. Store the reviewed catalog beside the digest so a mismatch is explainable. The gateway—not the model—fetches, compares, classifies, and enforces the result. The model receives only a compact status such as catalog_status=unchanged or catalog_status=review_required.
Copy this tool-surface contract
Apply this at initial connection, after a tool-change notification, before resuming a long-running task, and immediately before any write-capable operation.
1. DISCOVER AND RECORD
server_id: [pinned URL / trusted registry ID / certificate-bound identity]
retrieved_at: [UTC timestamp]
catalog: tools/list response across every page
canonicalization: [versioned deterministic procedure]
catalog_sha256: [server-client computed digest]
2. CLASSIFY THE SURFACE
For every tool, record:
- exact name and description
- input and output schemas
- declared annotations (informational, never proof of safety)
- risk class: read | reversible_write | external_effect | destructive | unknown
- allowed scopes, approval rule, and owner
3. BIND REVIEW TO THE FINGERPRINT
proposal_or_session_id: [stable ID]
approved_catalog_sha256: [exact digest]
allowed_tools: [small allowlist]
approved_until: [short expiry]
4. BEFORE A TOOL CALL
- Re-list after a change notification or reconnect.
- Recompute the fingerprint outside the model.
- Require exact digest match for an existing approval.
- Validate tool name and arguments against the reviewed schema and allowlist.
- If changed, return `tool_surface_changed`; do not select a “similar” tool.
5. ON A CHANGE
- Diff old and new catalogs by tool name, description, schema, and annotations.
- Mark added tools, removed tools, and changed parameters separately.
- Require new review for any material change or ambiguous classification.
- Log both fingerprints, the diff, reviewer/policy decision, and final call receipt.
For example, a tool called publish_report may retain its name while gaining a recipients argument. That is not a harmless catalog refresh: it changes the possible external effect. A digest mismatch should prevent the old approval from authorizing the new schema. Conversely, a changed marketing description for a read-only lookup may deserve a logged review but not the same escalation. Define this classification in code and policy, not in the server’s annotations or the model’s interpretation.
Handle the notification without trusting it blindly
A list_changed notification means “refresh the inventory,” not “the change is safe.” Re-run discovery across every paginated result, because a catalog that stops after page one is not a complete fingerprint. If the server does not advertise change notifications, choose a conservative refresh point: connection start, session resume, and before any consequential action. A short-lived agent task may only need one inventory; a persistent assistant needs a stated revalidation interval and a stop rule.
Failure mode: treating a hash as a safety verdict
A matching fingerprint cannot prove that the tool implementation is safe, that its result is trustworthy, or that a write is appropriate. A malicious or compromised server can preserve its schema while changing behavior. A hash also becomes noisy if descriptions contain timestamps or generated text, encouraging teams to ignore changes.
Use the fingerprint as a change-control guardrail alongside server trust, least-privilege credentials, argument validation, approval for consequential effects, and receipts that prove what actually happened. Keep the canonical form narrow enough to be stable, but include every field that changes what the model can call or how the application validates it. The point is not to make a changing tool catalog impossible. It is to make change visible before an agent inherits it.
The habit to keep
do not approve a server name and assume the tools stay the same. Approve a specific callable surface, bind it to a digest, and stop for fresh review when that surface changes.