Workhorse for AI coding agents

Workhorse gives an AI coding agent a direct path from documentation discovery to verified application code. Start with the narrowest useful Markdown source, then follow the same integration sequence in TypeScript, Python, or Go.

Read the documentation without an HTML shell

• Fetch /llms.txt for a compact map of every documentation page, grouped like the sidebar.
• Fetch /llms-full.txt when the task needs the complete documentation corpus in one response.
• Append .md to a page URL for its Markdown twin, such as /docs/enqueue.md.
• When the agent starts from HTML, follow the response's text/markdown alternate link to the same page's Markdown twin.

Use the per-page twin for focused work because it spends less context. Use /llms-full.txt when a change crosses several contracts and the agent needs to search them together.

Integrate one job end to end

Install the SDK used by the application:

• TypeScript: npm install @stablemates/workhorse
• Python: pip install stablemates-workhorse
• Go: go get github.com/stablemates/workhorse/go

Each complete example below opens PostgreSQL, enqueues order.created inside the transaction that inserts the order, runs a matching worker, and reads the durable result. Production processes call the worker's continuous run method; these bounded examples use one pass so they can finish.

import { Admin, Pool, Queue, Worker } from "@stablemates/workhorse";

const pool = new Pool({ connectionString: process.env.DATABASE_URL });
const queue = new Queue(pool);
const client = await pool.connect();
let jobId: string;

try {
  await client.query("BEGIN");
  await client.query("INSERT INTO orders (id, status) VALUES ($1, $2)", ["order-42", "new"]);
  jobId = await queue.enqueue("order.created", { orderId: "order-42" }, {}, client);
  await client.query("COMMIT");
} catch (error) {
  await client.query("ROLLBACK");
  throw error;
} finally {
  client.release();
}

const worker = new Worker(queue).handle("order.created", async (payload: { orderId: string }) => ({
  processedOrderId: payload.orderId,
}));
await worker.runOnce(); // Production worker processes call worker.run().

const job = await new Admin(pool).getJob(jobId);
console.log(job?.state, job?.result);
await pool.end();

The application transaction covers acceptance, so the order and job commit or roll back together. The handler runs later and may run again after a retry, so protect its external effects with an idempotency key, outbox, inbox, or compensation path.

Queue, Worker, and Admin are the TypeScript names, including Admin.getJob for confirmation. Python keeps the types and uses snake-case methods such as Admin.get_job; Go uses workhorse.NewQueue, workhorse.NewWorker, and workhorse.NewAdmin, with methods such as Admin.GetJob.

Next

• Agent workflows — compose durable model, tool, timer, and approval boundaries
• Enqueue a job — understand transaction and routing options
• Workers — run handlers continuously in production

Exact enqueue and transaction semantics: architecture reference.