dsh-agent-gate

When you run multiple Claw Agent sessions in parallel, an unauthorized MCP call or an over-privileged skill command can slip through the tool layer silently - by the time you notice in the logs, the damage is done. dsh-agent-gate inserts a layered permission gate at both the pre-execute and tools.guard stages, ensuring every invocation is checked against the actual policy.

Its capabilities go well beyond a simple yes-or-no block. It computes the intersection of official, Agent-level, and session-level permissions and applies them per call. Skills on the denylist are removed from the session skill catalog and command menu, not merely blocked at the skill tool level. Unauthorized MCP services are rejected and stripped from the model visible tool list at session start. When human review is required, the gate routes the request through the official one-time approval flow, and every event is recorded in audit.jsonl with a filterable dashboard that lets you slice by time, result, tool type, and Claw Agent, then export the current view. The official file sandbox for Claw sessions is pinned to read-only or workspace-write by the companion dsh-session-permissions plugin, and the gate can be uninstalled independently without loosening that baseline.

Ideal for teams operating multiple agents with explicit security requirements on tool invocations. If you already rely on dsh-session-permissions for permission boundaries, dsh-agent-gate is the missing runtime interception layer that closes the loop.