What is Rewind

Rewind is a runtime for programmatic tool calling, built for the developers of agent harnesses. Instead of calling your tools one at a time through the conversation, the model writes a TypeScript script that calls them: your MCP servers, your own functions, other workflows. Rewind type-checks the script against your tools, runs it in a deterministic V8 sandbox, and records every interaction with the outside world in a journal.

The model writes this

import { github, slack } from "rewind:tools";
import { sleep, signal } from "rewind";

export default async function main(input: { repo: string }) {
  const issues = await github.list_issues({ repo: input.repo, state: "open" });
  const stale = issues.filter((i) => Date.now() - Date.parse(i.updated_at) > 30 * 86_400_000);

  await slack.post_message({ channel: "#triage", text: `${stale.length} stale issues in ${input.repo}` });

  const decision = await signal<{ close: boolean }>("approval", { timeout: "3d" });
  if (!decision?.close) return { closed: 0 };

  await Promise.all(stale.map((i) => github.close_issue({ repo: input.repo, number: i.number })));

  await sleep("10d");
  const reopened = await github.list_issues({ repo: input.repo, state: "open", labels: "reopened" });
  return { closed: stale.length, reopened: reopened.length };
}

One script, one tool call from the model’s point of view. It lists the issues, filters them in code, waits three days for a human, closes the stale ones in parallel, sleeps ten days and checks again. Only the last line comes back into the model’s context.

Everything comes from the journal

Every tool call, timer, signal and log line is an entry in the run’s journal, linked to the line of TypeScript that caused it. The journal is the only source of truth: every feature of Rewind is derived from it.

Durable execution

await sleep("10d") or wait for a signal. The run survives crashes and deploys, and resumes by re-executing the script against its journal.

Replay

Any run re-executes from its journal with identical results. Fix a failed run with a fork: the calls it already made are not made again.

Testing

A recorded run becomes a cassette of mocks. Time is virtual, so a ten-day workflow is tested in milliseconds.

Observability

OpenTelemetry traces projected from the journal, and a dashboard of every run and all the code executed, with a debugger that steps through each run.

Who it is for

Rewind is for teams that build agents: a coding agent, a support agent, an operations agent, any product where a model calls tools for your users. Your harness keeps its model loop and its prompts. It adds Rewind through the TypeScript SDK (or the HTTP API /v1), and gives the model a few ready-made tools: run_script, get_run, signal_run, retry_run and others. Each of your users gets a namespace: their runs, scripts, credentials and quotas, isolated from everyone else’s.

Get started

How it works

Run Rewind

What is Rewind

Everything comes from the journal

Durable execution

Replay

Testing

Observability

Who it is for

Where to go next

Quickstart

Why Rewind

How it works

The dashboard