// find repeated agent work

Turn completed agent work into reusable know-how.

Punk learns from completed agent work and shows which tasks could safely reuse a proven result. New or uncertain work stays with your original model.

No credit card required. Your 30-day Pro trial begins after email verification. After the trial, keep Free (10,000 runs/month) or choose Pro ($99/month).

See a sample report Review security Compare plans

What happens next

Three setup steps get one compatible request into Punk. Run details close the loop, and observe-only is the default starting point.

01 / create

Create your account

02 / copy

Copy your Punk API key

Create a Punk API key in observe mode. It lets Punk record and explain runs, but not replace their live responses.

03 / connect

Change one base URL

Point a compatible OpenAI client at Punk, or use the Punk SDK. Your existing model and prompts stay in place.

Your first run, end to end

Send a familiar test request, then inspect the receipt Punk creates for it.

1 / CreateOpen your workspace and confirm your email.

2 / CopyCreate and copy a Punk API key in observe mode.

3 / ChangeUse Punk as the base URL for a compatible client.

4 / SendRun one representative request through your existing model.

5 / SeeOpen the run details to review route, cost, policy, and timing.

Developer checklist

1Confirm your email and open the dashboard.
2Create a Punk API key in observe mode under Governance → Punk API keys.
3Set baseUrl to https://app.punktechnologies.com in the SDK, or point a compatible OpenAI client at https://app.punktechnologies.com/v1.
4Send a test request and open its run details. Repeated representative traffic gives Punk more evidence to analyze.

Install

npm install @punktechnologies/sdk

import { Punk } from "@punktechnologies/sdk";

const punk = new Punk({
  baseUrl: "https://app.punktechnologies.com",
  apiKey: process.env.PUNK_API_KEY,
  app: "support-agent"
});

Read setup docs

See the first run for yourself

Connect without changing live answers. Punk first finds repeated work; you decide what is worth testing for safe reuse.

Create account See a real example

Planning a focused pilot? Join the design-partner program