Every team I speak with building Claude agents eventually encounters a common hurdle: SOC 2 auditors lack the vocabulary to understand “this agent utilizes fourteen MCP tools spanning Stripe, Salesforce, Slack, and Gmail.” Compliance reviews still occur at the project level, annually, typically after the agent has been operational in production for months. There’s no equivalent of a linter for the crucial question when configuring an agent: what regulatory exposure am I assuming by integrating these specific tools? What I appreciate about SCOPE is its deliberate avoidance of the LLM-as-a-judge approach. The risk levels and regime tags (GDPR, HIPAA, PCI, SOX, SOC 2, EU AI Act, ISO 42001, NY DFS 500—26 codes in total) originate from a curated database encompassing over 80+ MCP servers. This consistent input-output process ensures determinism, which is essential for the output to seamlessly integrate into a real change-management workflow rather than becoming another “trust me” AI artifact. Furthermore, the entire dataset is publicly accessible, allowing users to open a pull request against the YAML file if they disagree with how slack.read_direct_messages is tagged for their specific context. SCOPE operates at the design phase as a Claude plugin, running before the agent is deployed. This allows for the modification of the toolset during the development process, as changing tools is relatively inexpensive. This flexibility includes options like dropping a tool, swapping a write operation for a read, or requiring human approval for critical actions. SCOPE seamlessly integrates with existing runtime guardrails. I’m thrilled to have the LangGuard team here today to delve into the data model, regime coverage, and the relationship between pre-flight evaluation and runtime enforcement. Congratulations on the successful launch!

SCOPE MCP

A free linter for agent permissions, before they ship.

A free linter for agent permissions, before they ship.