Agentic Use is where you decide what your coding agents are allowed to do — which shell commands they may run and which MCP tools they may call. Policies in the sidebar expands to its groups; Agentic Use is the first. Opening
/policies directly lands here too.
What’s on the page
Tab strip
Policy list
Policy Health
Shows how well your tool policies cover what your agents actually run. A Terminal / MCP switch picks which kind of policy to read, and 7 days · 30 days · 90 days in the header sets the window. Four cards sum it up:
Overview lays the findings out in three columns:
Recently Edited lists policies changed in the window with when each was created and last edited. A policy still collecting data is badged Evaluating, with a note such as Metrics will stabilize in 5 days, so a fresh edit is not judged before it has had traffic.
Policy Health needs Policies · View.
Environment Targets
Every host, database, cloud resource, secret name, container image, repository, branch, file path and other target your agents have touched. Each row carries the command family it came from, how many developers used it, when it was last seen, how many of those uses were high risk, and the total number of times it has been seen. You tag each one Low, Med or High sensitivity. From then on, commands against that target are scored in the band you named rather than the one Unbound would have guessed — your own knowledge of which database is production wins. Commands already classified keep the score they were given. Untagged paths fall back to defaults:/etc, /usr, /var, /boot, /sys, ~/.ssh, ~/.aws, ~/.gnupg and .env files read as High, and ~/.config, .github, Dockerfile and docker-compose* as Med.
Filter by command family or by sensitivity (including Untagged, which is the working list), and search for a target by name.
Filters and controls
Search and filter the policy list by name, type and status. Group scoping is set per policy, drawn from Policies → Users → Groups.What you can do
A policy’s own page
Match panel
The policy types
The picker asks What do you want to monitor? and each choice carries a line of its own — Control which shell commands AI coding tools can execute, Control which MCP tools AI coding agents can use, and for Skills, Require AI coding agents to invoke specific skills when policies match. Ask AI asks the same question with three choices — Terminal commands, MCP actions and Skills — then drafts policies from your description for you to review. The picker offers Terminal Commands, MCP Actions and Skills.
A skill policy opens at
/policies/skill-policies/[id]. Probe policies are created from Create Policy → MCP Probe.
Ask AI
/policies/agentic-use/ask-ai drafts a policy from a plain-language description for you to review.