Engineering briefs, architecture deep-dives, and product updates from the team building the trust layer for AI agents.
Claude can ask for any tool you give it. governed_tools checks each call first, and a refused call never runs.
Claude agent guardrails for tool use: governed_tools checks every call before it runs. add is approved, an undeclared divide is blocked.
Your MCP server runs whatever a client calls. MCPServerGuard checks each call first, and a refused call never runs.
MCP server security in one line: MCPServerGuard checks every tool call before it runs. add is approved, an undeclared divide is refused.
governed_functions wraps the functions you pass to register_function. A refused call never runs.
Block unsafe AG2 tool calls before they run. governed_functions checks each call: add is approved, an undeclared divide is blocked.
governed_tools wraps the list you pass to AssistantAgent. A refused call never runs.
Block unsafe Microsoft AutoGen tool calls before they run. governed_tools checks each call: add is approved, an undeclared divide is blocked.
Put @harness on the function that has the side effect. A $40 refund runs. A $5,000 refund does not.
Stop risky AI agent tool calls in Python. The @harness decorator checks each call before the function runs: a $40 refund runs, a $5,000 refund is blocked.
governed_tools wraps the list you pass to bind_tools and ToolNode. AgentTrustNode is not that check.
Block unsafe LangGraph tool calls before they run. governed_tools checks each call at the tool edge: add is approved, an undeclared divide is blocked.
New engineering briefs and product updates whenever we ship something worth reading.