MCP or Claude Skills for Approvals?
Use MCP to retrieve live approval records and a Claude Skill to apply a reusable review procedure. When a workflow needs both, test the record lookup and the decision separately so a connection failure cannot pass as a policy result.
Put the clerk’s procedure in the skill
Use a Claude Skill when the agent already receives the information it needs and must apply the same procedure each time. For an approval review, put the required fields, threshold rules, and handling of incomplete records in the skill. Bundle a template or reference file when the procedure needs one. Give the skill a description that names the record type and the intended review.
Test the procedure with a complete record, a record at the decision threshold, and a record missing a required field. A passing run gives the expected disposition and identifies missing information before reaching a decision. If the procedure works when explicitly requested but is skipped in ordinary requests, revise the description and repeat the discovery test.
Use a replay in which the first execution succeeds and a retry arrives with the same action ID. The second execution should be observable even if an operator sees only one final record.
Equipping agents with Agent Skills \ AnthropicPut live record access behind MCP
Use MCP when the agent needs to retrieve a record from another system as part of the task. The application acts as the host and coordinates clients, each of which communicates directly with a server. Start with a read-only lookup that accepts a record ID and returns the fields the review uses. Keep access checks at the service that owns the data.
Test with one accessible fixture, one missing ID, and one record the test identity cannot read. Inspect the tool result for each case before adding the review skill. The workflow should preserve the difference between a successful retrieval, a missing record, and a denied request. If those outcomes become indistinguishable, fix the lookup contract before relying on the agent’s decision.
Keep the policy decision, the human response and the tool invocation as distinct events. A recorded request for approval by itself does not show that the user approved the action.
Understanding MCP clients - Model Context ProtocolTest the handoff between lookup and judgment
In an illustrative trial, request P-104 is for $1,250 and has no approver. The policy requires a named approver above $1,000. Have the MCP lookup return the ID, amount, and empty approver field. The skill should report that the request needs an approver. Then return the same record with an approver and check that the report identifies the policy requirement as satisfied.
Repeat with the lookup returning an access denial. The response should report that the request could not be reviewed and omit any policy disposition. Record the returned fields and the final outcome separately. If the retrieval is correct but the disposition is wrong, review the skill instructions. If the returned record is incomplete, inspect the lookup and its response mapping.
Agent Skills - Claude Platform DocsChoose the transport for the operating environment
Choose stdio when the MCP client should launch the server as a subprocess in the same operating environment. Choose Streamable HTTP when you need an independently running service that can accept connections from multiple clients. For stdio, check that stdout carries MCP messages and direct diagnostic logs to stderr. For Streamable HTTP, verify Origin validation and test access with the identities you intend to support.
Run the same read-only fixture from two intended clients before choosing a shared service. Check that each receives the fields its identity is permitted to see. With stdio, a process that starts and immediately loses the connection calls for inspecting stdout and the subprocess logs. With HTTP, inspect the service response and the client’s access configuration before changing the skill.
Transports - Model Context Protocol