Scope down what Claude Code and Cowork can access, and check every tool call against your policy before it runs.
Claude Code and Cowork act with the full access of the person behind them, often connected to company apps and MCP servers without a security review. A valid login doesn’t settle whether an action should run, and a mistaken or misled agent can take any action its user could.
This brief shows how Obsidian closes that gap in two steps. First, it finds each Claude agent’s MCP servers, skills, and plugins, and cuts access the work doesn’t need. Then a sidecar checks each tool call against your policy before it reaches the app, using identity, HR, and application context along with your own data labels. Every decision is logged with the agent, user, tool, arguments, and matching policy.
Read it to see how the decision flow works, how a contractor’s agent is stopped from dropping a critical Snowflake table, and four questions to ask any vendor about agent security.