The agent proposes. PRAE decides.

Control what autonomous AI agents can execute. PRAE evaluates actions against runtime policies, enforces decisions at supported execution boundaries, and records verifiable evidence.

What PRAE controls

  • Every tool call that passes Claude Code’s PreToolUse hook: built-in tools and MCP tools.
  • Four outcomes: allow, deny, ask (Claude Code’s permission prompt), and clamp (refused, with the permitted value named).
  • A signed, hash-chained ledger row for every decision, with the clause that decided it.

What it does not control

  • What the model writes. PRAE judges actions, not text, and detects no prompt injection.
  • What a shell command does once it runs. Bash is judged by its command line, not its effect.
  • Anything that does not pass the hook, such as commands you type after ! yourself.

The supported evaluation path

  • This page: the policy kernel, in your browser.
  • Then one Claude Code laptop: npm install -g prae-cli, prae init claude-code, shadow mode first (developer agents).
  • The limits are in the evidence package.

See how PRAE evaluates agent actions, applies the policy decision, and records the outcome: four Claude Code tool calls against pack prae-evaluation 1.0.0, three rules you can read in full. Trigger each one, read the decision, then verify and export the ledger at the bottom.

Simulated boundaryThe real policy kernel decides each call in your browser, with the Claude Code hook’s mapping. No agent runs, no file is read and nothing leaves this page until you download it.
Trigger a scenario to see what the boundary decides, why, how long it took, and the row it leaves.

Ready to test PRAE against your own agent environment? Book a short kickoff with our team: we learn your setup, define the evaluation scope, and help launch your dedicated sandbox.

Start a guided evaluation →
  1. Download the evidence, the ledger and the pack above.
  2. npx prae-cli verify prae-<session>.ledger.jsonl recomputes every hash and confirms each row names the one before it. It prints chain intact and the seal that matches the page.
  3. Download the tampered ledger instead: the same command prints CHAIN BROKEN and the line that no longer matches.
  4. npx prae-cli explain evaluation.prae.json Read file_path=/home/dev/.ssh/id_ed25519 re-runs a decision against the pack, independently of the ledger.
  5. shasum -a 256 evaluation.prae.json equals the pack hash every row carries.

An intact chain is not proof that a decision was right. It shows the rows were not edited, removed or reordered after they were written. Whether each decision follows from the policy is what step 4 checks. These browser rows are unsigned; on a real install every row is also signed with an Ed25519 key you hold, and prae verify --key checks those signatures.

The same four calls, measured through the real Claude Code hook with a signed ledger. Run on 2026-10-27, commit de0ad56201eb: darwin 25.5.0, Apple M5 Max, Node v26.5.0, prae-cli 0.10.6.

Call Policy decision (kernel) Gate, in process Hook process, as Claude Code runs it
Allow p50 1 µs · p99 1.5 µs p50 0.47 ms · p99 0.95 ms p50 43 ms · p95 49.3 ms
Deny p50 1.7 µs · p99 2.5 µs p50 0.47 ms · p99 0.94 ms p50 42.6 ms · p95 44.3 ms
Ask p50 0.6 µs · p99 0.8 µs p50 0.53 ms · p99 1.06 ms p50 42.2 ms · p95 45.7 ms
Clamp p50 0.1 µs · p99 0.2 µs p50 0.44 ms · p99 0.74 ms p50 44.1 ms · p95 47.7 ms
  • Gate, in process: loading the pack, judging the call and appending a signed ledger row, once Node is running (1000 runs each).
  • Hook process: one prae hook claude-code process per call, payload on stdin, from the built npm package (30 runs each). Starting Node is most of it.
  • Verifying 1,000 signed rows: 52.8 ms in process, 110.5 ms as a prae verify --key command.
  • Verifying 10,000 signed rows: 524.4 ms in process, 608.2 ms as a prae verify --key command.

Time one call on your own machine (after npm install -g prae-cli and downloading the pack):

echo '{"hook_event_name":"PreToolUse","session_id":"try","cwd":"/tmp/acme","tool_name":"Read","tool_input":{"file_path":"/tmp/acme/.ssh/id_ed25519"}}' > call.json
time prae hook claude-code --enforce --pack evaluation.prae.json --ledger try.jsonl < call.json

Don't just evaluate the boundary. Test it in your environment.

You've seen how PRAE evaluates, controls, and records agent execution. Now bring your own agent environment: book a short kickoff with our team, define your evaluation scope, and launch a dedicated PRAE sandbox together.

25-minute kickoff · Guided integration · Dedicated evaluation sandbox

Also: the evidence package · trust center