quickstart

Lelu quickstart — see it work in 60 seconds

This runs the real engine on your machine — SQLite only, no Postgres, no Redis, no cloud account — and fires one request for each of Lelu's four authorization outcomes, including a live prompt-injection catch.

Run it

git clone https://github.com/lelu-ai/lelu
cd lelu/examples/quickstart
./demo.sh

Requires Go 1.24+, curl, and python3 — nothing else. The script builds the engine, starts it, runs the requests, and cleans up after itself.

Already have the engine built? Skip the build: LELU_ENGINE_BIN=/path/to/lelu-engine ./demo.sh

What you'll see

Each call goes to POST /v1/agent/authorize. The engine returns boolean flags (allowed, requires_human_review, compute) — your SDK turns those into one of four decisions. This is actual output from the run:

1. ALLOW — a permitted action with high confidence:

{ "allowed": true, "requires_human_review": false, "reason": "action authorized" }

2. DENY — a prompt injection hidden in the request, caught before policy even runs:

{
  "allowed": false,
  "requires_human_review": false,
  "reason": "prompt injection detected in resource: \"ignore all previous\""
}

3. HUMAN_REVIEW — the model wasn't confident enough, so it's held for a human:

{
  "allowed": false,
  "requires_human_review": true,
  "reason": "confidence 80% requires human approval (threshold 90%)"
}

4. COMPUTE — a risky production deploy is redirected to a safe sandbox instead of blocked:

{
  "allowed": true,
  "requires_human_review": false,
  "compute": true,
  "safe_tool": "deploy",
  "reason": "action \"deploy_app\" redirected to safe alternative \"deploy\" by compute rule"
}

The policy behind it

All four behaviours come from one small file, policy.yaml:

• invoice_bot may approve_refunds and deploy_app, never delete_invoices.
• Confidence gating: >= 0.90 allow · 0.70–0.89 review · 0.50–0.69 read-only · < 0.50 deny.
• A compute rule rewrites any deploy_* into a sandboxed staging deploy.

Edit policy.yaml, re-run ./demo.sh, and watch the decisions change.

Next

• Call it from your app with the TypeScript or Python SDK.
• Run the full stack (dashboard, Postgres, Redis) with docker compose up from the repo root.