MCP Tool Evaluation with Claude Code: Sandbox Policy

A production playbook for MCP tool evaluation in cross-industry operations using Claude Code: sandbox policy, run-scoped inputs, logs, typed results, and artifacts.

Audience: AI platform teams adopting MCP

The problem

AI platform teams adopting MCP need MCP tool evaluation to run repeatedly against tool definitions, auth policy, traces, and test cases. In cross-industry operations, the pain is not one good answer; it is repeatability, auditability, exception handling, and evidence that survives handoff.

Implementation path

Run MCP tool evaluation in an ephemeral sandbox, keep provider credentials in the broker, expose narrow tools, and store logs outside the workspace for review.

Tradeoffs and failure modes

A narrower runtime blocks ambient machine behavior, but it gives security reviewers a concrete boundary. For MCP tool evaluation, the practical test is whether a second run can be debugged, retried, and consumed by a product without reading the raw agent transcript.

Runtime boundary

filesystem: /skill and /skill/.argo/inputs only
network: deny by default
artifacts: /skill/output/artifacts
logs: retained outside sandbox
provider: Claude Code

Run this on Argo