PRAE for Claude Code, on every developer machine.

Shipping

Claude Code reads files, runs commands and calls MCP tools with a developer’s access. PRAE installs a PreToolUse hook in its managed settings, for every user on the machine, so each built-in and MCP tool call is judged against your policy before it runs. Across a fleet, the control plane enrolls each machine and keeps its policy current.

Serves: Developer agents →

  1. 01 · Where the agent operates

    Claude Code

  2. 02 · What it can reach

    The repository and files on the developer’s machine, the shell, the web, and the MCP servers it is configured with.

  3. 03 · Where PRAE intervenes

    Before each tool call runs: Claude Code hands the call to the hook, and runs it only if the hook allows it.

04 · What PRAE does: Evaluate → Enforce → Steer → Prove

  1. Evaluate

    Each built-in and MCP tool call is judged against the policy in force: tool, arguments, paths, hosts, commands and budgets.

  2. Enforce

    A denied call never runs and Claude Code is told why; a held call becomes Claude Code’s own permission prompt. With Claude Code 2.1.295 or later, a hook that fails to run blocks the call instead of letting it through.

  3. Steer

    On an enrolled machine, a policy promoted, rolled back or frozen in the control plane applies from the next tool call, with no restart.

  4. Prove

    Every decision is sealed in a signed, hash-chained ledger for each user, and shipped to the control plane as metadata.

05 · What you get, and how it plugs in

  1. One machine npm install -g prae-cli && prae init claude-code --scope user --enforce --apply
  2. A fleet: one script per OS, from a registration code prae org install-script --file platform.json --os macos --out install-prae.sh

Every tool call Claude Code routes through its hooks: built-in tools and MCP tools, in the CLI, the IDE extensions and the desktop app, which read the same managed settings.

A shell command is judged by its text, not its effect; files attached with @ fire no hook; tool output is not screened; what a person answers to a hold is not recorded. A local administrator can change managed settings: the change is detected and reported, not prevented. Claude Code on the web does not read on-device settings.