Getting Started

Record any workflow on any app with our browser extension (Chrome or Firefox). Trayn creates a fully interactive, anonymized RL playground where agents can safely practice sending emails, escalating tickets, placing orders, and every other operation you'd never risk on a live system.

Each action is graded in real-time and packaged into training memories the agent retrieves to improve on the next attempt.

How It Works

Architecture

The training loop runs through both packages:

1. @trayn/agent-sdk fetches the task from your playground, drives a real browser, and sends each step to the grader for multimodal evaluation
2. @trayn/memory stores graded memories with environment embeddings, then retrieves relevant REPEAT/AVOID memories on subsequent reps so the agent learns from mistakes

Installation

npm install @trayn/agent-sdk

bun add @trayn/agent-sdk

pnpm add @trayn/agent-sdk

yarn add @trayn/agent-sdk

The SDK requires Playwright. Install browsers after:

npx playwright install chromium

Installing @trayn/agent-sdk pulls in @trayn/memory automatically.

Quick Start

trayn --url https://app.trayn.ai/playground/{host}/{sessionId} --max-steps 15 --reps 3

import { harness } from "@trayn/agent-sdk";

await harness({
  agentargs: {
    agent_name: "my-agent",
    make_agent: () => ({
      get_action: async (obs) => {
        return ["click('42')", { think: "clicking submit" }];
      },
      store_grades: async (result) => {
        console.log("Completed:", result.taskCompleted);
      },
    }),
  },
  url_override: "https://app.trayn.ai/playground/{host}/{sessionId}",
}).run();

Training Outputs

Every training run produces structured data you can consume directly or feed into other platforms:

Memories use a REPEAT/AVOID format — successful steps become REPEAT memories, failed steps become AVOID memories with a correction field describing what the agent should have done instead. On subsequent repetitions, relevant memories are retrieved and injected into the agent's prompt:

action -> FAILED: clicked wrong button | DO INSTEAD: click the "Save" button in the toolbar

This data is portable. Export memories and grades from the Experiments panel or read them via the @trayn/memory package to plug into your own fine-tuning, RLHF, or evaluation pipelines.

Authentication

To access your playgrounds from the CLI or SDK, you need to authenticate with your Trayn account. This is the same account you use to sign in at app.trayn.ai.

On first run, the CLI prompts for your email and saves it locally:

trayn --url https://app.trayn.ai/playground/...
#  Email: you@company.com (saved to ~/.config/trayn/config.json)
#  ✓ Saved

This credential links all your training runs to your account — grading results, memories, and experiments all appear under your profile in the web app. It also authorizes access to your playgrounds so only you (and your team) can run agents against them.

CLI Reference

trayn [options]

Programmatic Usage

import { harness } from "@trayn/agent-sdk";

await harness({
  agentargs: {
    agent_name: "my-agent",
    make_agent: () => ({
      get_action: async (obs) => {
        // obs.goal — task objective fetched from the backend
        // obs.url — current playground page URL
        // obs.axtree_txt — accessibility tree with element BIDs
        // obs.last_action_error — feedback from previous action
        return ["click('42')", { think: "clicking the submit button" }];
      },
      store_grades: async (result) => {
        console.log("Completed:", result.taskCompleted);
      },
    }),
  },
  url_override: "https://app.trayn.ai/playground/{host}/{sessionId}",
}).run();

Environment Variables

The CLI and SDK need an API key for whichever AI provider your agent uses. Trayn supports OpenAI, Anthropic, Google, and any provider compatible with the Vercel AI SDK.