#What the agent may run

hush install-skill            # this project
hush install-skill --global   # every project

Installs a skill telling your agent never to ask for a pasted key, to use hush_run rather than reading values, and how to pick a set when you name one. It goes where each agent looks for instructions — the right-hand column of the table in Connecting your agent. Codex, Gemini CLI and Zed share .agents/skills/ (and ~/.agents/skills/ for --global), so one file serves all three. Without it the tools still work — the tool descriptions explain themselves — you just have to say so each time.

.hush/policy.json controls what it may run — through the MCP tools and through the CLI, so an agent that shells out to hush run or hush export meets the same approval prompt. For the agent's tools, anything that exists to dump or re-encode the environment is denied outright — shells, env, base64, curl, and every interpreter, because node -e can write the whole environment to a file that output redaction never sees. In your terminal the same commands are not refused: hush node server.js goes to the approval prompt with a warning line, the way op run asks rather than blocks. allowCommands still narrows both:

{
  "allowCommands": [],
  "denyCommands": ["your-own-additions"],
  "unsafeAllowCommands": [],
  "allowEnvs": ["default", "work-fal"],
  "allowHosts": ["api.stripe.com", "*.github.com"],
  "denyKeys": ["STRIPE_LIVE_KEY"],
  "maxRunMs": 120000,
  "approvalScope": "command"
}

• denyCommands is added to the built-in list, never substituted for it — an old config cannot hold you below the current floor.
• unsafeAllowCommands is the only way below that floor. Named to be off-putting: allowing node or bash lets an agent read every injected secret and write it anywhere.
• allowEnvs names the sets an agent may use — by the name you gave them, wherever they live — so ["default", "work-fal"] keeps an agent out of client-fal.
• allowHosts bounds where hush request and hush_request may send a credential. Empty means any host over https. Entries are a bare host (api.stripe.com), a host with a port (localhost:3000), or a subdomain glob (*.example.com, which does not match the apex). It is the one control that decides where a secret goes, which no command or set rule can express: without it, a set allowed for api.stripe.com is equally allowed for an attacker's collector.
• allowCommands, if non-empty, is an allowlist — the only command control that actually holds. Prefer it for anything sensitive.
• requireApproval lists what a human must approve on screen: run, add, reveal, and request. --materialize is gated by reveal, not run, because it writes plaintext to a path the caller chose.
• approvalScope is what "Allow 15 min" covers: "command" (the default) means that command with those sets; "sets" means any allowed command with those sets, if you find the prompts too frequent.
• approvalTtlSeconds is how long an "Allow" lasts (default 900 = 15 minutes), counted inside the process that asked — the MCP server and the app keep it, a one-shot hush command does not. Change it in the app's Agent section, or with hush secure approval --for 30m, rather than by hand.

Your floor. policy.json is a file in the repo, so an agent with write access can edit it. ~/.hush/policy.json — same shape, outside every repo — is your floor: a repo's policy can only tighten relative to it. allowCommands, allowEnvs and allowHosts can only narrow, requireApproval and denyKeys can only grow, biometry and approvalScope can only get stricter, and unsafeAllowCommands only takes effect when your floor lists the same command — the repo asks, you permit. unmaskKeys is floor-only as well: it is the list of keys you have allowed to print unmasked, and a repository cannot add to it. hush doctor shows every place a repo policy tried to go below your floor.

An allow list your floor sets keeps applying even when a repo's policy.json does not mention it. A repo file that omits allowCommands inherits your floor's list rather than replacing it with "no restriction"; a repo file that names only entries outside your floor is ignored as a whole. Both cases are reported by hush doctor, so a floor that is doing nothing is never silent.