AgentPerms

AgentPerms is a small agent.permissions.yaml schema, validator, and evaluator for AI-agent permissions.

It defines what an agent may:

• call as tools
• read and write as files
• use as models
• spend per run
• expose to external destinations

Install

corepack pnpm install
corepack pnpm build
node dist/cli.js --help

CLI

node dist/cli.js init
node dist/cli.js validate agent.permissions.yaml
node dist/cli.js eval agent.permissions.yaml --tool github.pull_request.comment
node dist/cli.js eval agent.permissions.yaml --file .env --file-mode read

Blocked evaluations exit with code 2, which makes the CLI useful in CI and runtime wrappers.

Scope

AgentPerms is a manifest and policy evaluation layer. It does not enforce anything by itself. Pair it with MCPGuard, BrowserGuard, AgentPassport, or another runtime that actually blocks unsafe actions.

Verification

corepack pnpm install
corepack pnpm format:check
corepack pnpm check
npm pack --dry-run