Write your agent with the AI SDK, Mastra, the OpenAI Agents SDK, LangGraph or plain TypeScript. agent-unit makes it durable without a rewrite, so it can pause for a human, sleep for a day, survive a crash and resume where it stopped, and deploys it to Node, Bun, Deno, Cloudflare, Vercel, Netlify or AWS Lambda. Every agent gets the same HTTP API, a resumable event stream, an MCP endpoint and an A2A agent card.

agents/refund.ts ──► agent-unit build --preset vercel ──► .vercel/output
       │                                                    ├─ POST /agents/refund/runs
  AI SDK · Mastra · OpenAI Agents · LangGraph · plain TS    ├─ GET  /runs/:id/events   (SSE, resumable)
                                                            ├─ POST /runs/:id/resume   (answer a pause)
                                                            ├─ POST /mcp               (every agent is a tool)
                                                            └─ GET  /.well-known/agent.json

It is to agents what Nitro is to servers, and it is built on Nitro: one build and one runtime contract, with presets for every host.

Durable without a rewrite. Model calls and tool calls are journaled automatically. When a run continues after a pause, a restart or a crash, completed work replays from the journal instead of running again, so the provider is not billed twice and a refund is not charged twice.
Pauses that cost nothing. run.interrupt() parks a run until someone answers, and run.sleep("2d") until a time passes. A parked run uses no compute, on any host.
One contract for every framework. The same run API, AG-UI events, manifest, MCP and A2A surface, whichever framework an agent uses, so a UI or an approvals inbox works with all of them.
No deployers to maintain. Framework authors write one small adapter; agent-unit and Nitro handle hosts, storage and scheduling.

Quick start

npm install agent-unit

// agents/refund.ts
import { defineAgent } from "agent-unit";

export default defineAgent({
  description: "Refunds an order after a human approves it.",
  async run(input, run) {
    const order = await run.step("load-order", () => loadOrder(input.orderId));
    const decision = await run.interrupt<{ approved: boolean }>("approve-refund", order);
    if (!decision.approved) return "declined";
    await run.step("charge", () => refund(order)); // journaled: resuming never repeats it
    return "refunded";
  },
});

npx agent-unit dev          # http://localhost:3000, reloads on change
npx agent-unit build        # .output/server/index.mjs for Node
node .output/server/index.mjs

curl -X POST localhost:3000/agents/refund/runs \
  -H 'content-type: application/json' -H 'accept: text/event-stream' \
  -d '{"input":{"orderId":"o_1"}}'
# … event: RUN_INTERRUPTED  {"interrupt":{"name":"approve-refund","payload":{…}}}

curl -X POST localhost:3000/runs/run_…/resume \
  -H 'content-type: application/json' -d '{"answer":{"approved":true}}'

Every file in agents/ is an agent named after the file (agents/support/index.ts serves as support). Export a framework agent or a defineAgent agent as the default export.

Bring your framework

agent-unit recognises agents from these frameworks when your package.json depends on them:

Framework	Export	What becomes durable	How it pauses
AI SDK	`new ToolLoopAgent(…)` or `streamText` settings	Every model call and tool call	`useRun().interrupt()` in a tool
Mastra	`new Agent(…)`	Every model call and tool call (on a fork; your agent is untouched); memory uses the run's thread	Tool approvals (`requireApproval`, `requireToolApproval`), `suspend()` in a tool
OpenAI Agents SDK	`new Agent(…)`	Model responses and function tools, per turn; handoffs included, `handoff()` options too	Tool approvals (`needsApproval: true`)
LangGraph	`graph.compile()`	LangGraph checkpoints, stored in agent-unit storage	`interrupt()` in a node, resumed with `Command({ resume })`
Plain TypeScript	`defineAgent(…)`	`run.step(…)`	`run.interrupt()`, `run.sleep()`

Every framework is tested on every host below (Node, Bun, Deno, Cloudflare Workers with Durable Objects, Vercel, Netlify and AWS Lambda): a run pauses for approval, the host restarts, and the resume repeats no model call or side effect.

Events follow AG-UI: text streams as TEXT_MESSAGE_*, a model's reasoning as REASONING_*, and RUN_FINISHED carries token usage per provider and model (also on the run as usage). An agent that asks for structured output (AI SDK output, Mastra structuredOutput, OpenAI Agents outputType) finishes with that object as its output.

A framework's own pauses become agent-unit interrupts, answered with POST /runs/:id/resume:

Tool approvals (OpenAI Agents needsApproval, Mastra requireApproval or requireToolApproval) pause as tool-approval with { callId, tool, arguments, agent }. Answer true, false or { approved }; a declined call is not run and the model is told so.
Mastra suspend(payload) in a tool pauses under the tool's name with that payload. The answer becomes the tool's resumeData when it runs again, as with Mastra's own resume.
LangGraph interrupt(value) in a node pauses under the node's name; the answer resumes it.

// agents/support.ts: an AI SDK agent, unchanged except for the approval
import { useRun } from "agent-unit";
import { openai } from "@ai-sdk/openai";
import { ToolLoopAgent, tool } from "ai";
import { z } from "zod";

export default new ToolLoopAgent({
  model: openai("gpt-5"),
  tools: {
    refund: tool({
      description: "Refund an order",
      inputSchema: z.object({ orderId: z.string() }),
      execute: async ({ orderId }) => {
        const { approved } = await useRun().interrupt<{ approved: boolean }>("approve-refund", { orderId });
        return approved ? await refund(orderId) : "declined";
      },
    }),
  },
});

When the run resumes, the model's first response comes from the journal, the tool call continues with the answer, and only new work reaches the provider.

Run primitives

Inside any run, useRun() (or the run argument of defineAgent) gives you:

Primitive
`run.step(name, fn)`	Run `fn` once; replays return the journaled result. `fn` receives `{ idempotencyKey }`
`run.idempotencyKey()`	The key of the step or tool call running now
`run.interrupt(name, payload)`	Park until `POST /runs/:id/resume` answers, then return the answer
`run.sleep("10m")`	Park until the time passes
`run.state.get/set/delete(key, { scope })`	State scoped to the `thread` (default), the `agent` or the `app`
`run.secrets.get(name)`	Read a secret from the host environment (Workers bindings included)
`run.emit(name, value)`	Emit a `CUSTOM` event
`run.signal`	Aborts when the run is cancelled or parked

The one rule: between steps, code must make the same decisions given the same journaled results. Put anything with side effects or randomness in a step. Model and tool calls already are.

A step is recorded when it finishes. If the process dies while a step is still running, that step runs again when the run recovers. Every step gets an idempotency key that stays the same when it runs again and differs for every other step and run; hand it to the API the step calls, and the repeat has no second effect:

await run.step("charge", ({ idempotencyKey }) =>
  stripe.refunds.create({ payment_intent: order.paymentId }, { idempotencyKey }),
);

Inside a tool the framework calls (AI SDK, Mastra, OpenAI Agents), useRun().idempotencyKey() returns the tool call's key. In a LangGraph node, wrap the side effect in useRun().step(…).

Deploy anywhere

npx agent-unit build --preset <preset>

Host	Preset	Storage	Long runs	Sleep and recovery
Node	`node-server` (default)	Filesystem by default	Unlimited	In-process timers + sweep every minute
Bun	`bun`	Filesystem by default	Unlimited	In-process timers + sweep
Deno	`deno-server`	Filesystem by default	Unlimited	In-process timers + sweep
Cloudflare Workers	`cloudflare-module` with `runtime: "durable-objects"`	One Durable Object per run (no storage to configure)	Yields every 10 min and continues from an alarm	Each run's own alarm
Cloudflare Workers (KV)	`cloudflare-module`	Set `storage` (for example `cloudflare-kv-binding`)	Yields every 25s and continues	Cron Trigger runs the sweep
Vercel	`vercel`	Set `storage` (Redis, Upstash, Vercel KV, …)	Yields every 240s and continues	Vercel Cron runs the sweep
Netlify	`netlify`	Set `storage` (`netlify-blobs`, Redis, …)	Yields every 20s; finishes in the request	Generated scheduled function
AWS Lambda	`aws-lambda`	Set `storage` (Redis, Upstash, a database via db0, …)	Yields every 25s; finishes in the request	Point EventBridge at `POST /__agent-unit/sweep`

Cloudflare builds turn on Workers' stub modules for child_process, readline, worker_threads and tty, and replace the Node ws package with the built-in WebSocket: Mastra and the OpenAI Agents SDK import these when they load, and one missing module stops the whole Worker from starting. Your own compatibility_flags and alias settings are kept.

Any other Nitro preset works too. Serverless hosts need shared storage, because their filesystem does not survive between invocations; the build warns when one is missing. Set AGENT_UNIT_SECRET in production: it protects the internal continue and sweep endpoints, which serverless hosts use to continue long runs. A long run continues in a fresh invocation at the deployment's own URL, read from AGENT_UNIT_URL (or VERCEL_URL, Netlify's URL, or the origin option) and never from an incoming request; without one it continues in-process.

Cloudflare Durable Objects

On Cloudflare, give every run its own Durable Object, the same primitive Cloudflare's own agents run on:

// agent-unit.config.ts
export default defineConfig({ preset: "cloudflare-module", runtime: "durable-objects" });

npx agent-unit build && npx wrangler deploy

Your agent code does not change. Each run's object holds its journal and events in strongly consistent storage and is the only thing that ever executes it, so there are no leases to race. Its alarm wakes it from run.sleep, continues it after a long step, and picks it up again if the isolate running it dies. A shared index object keeps the run list and thread, agent and app state. The build exports both classes and writes their bindings and migration into wrangler.json.

Your app can bring its own Cloudflare resources beside agent-unit's: D1, KV or R2 for a storage provider, or its own Durable Objects (Mastra's Cloudflare store ships one, as does a Workflow). Write them in Cloudflare's own wrangler.json format under cloudflare, and export your classes from a file:

export default defineConfig({
  preset: "cloudflare-module",
  runtime: "durable-objects",
  cloudflare: {
    wrangler: {
      d1_databases: [{ binding: "DB", database_name: "memory", database_id: "…" }],
      durable_objects: { bindings: [{ name: "STORE", class_name: "StoreObject" }] },
    },
    exports: "./store.ts", // or exports.cloudflare.ts, picked up on its own
  },
});

To connect your storage the same way on every host, read your bindings and secrets from agent-unit/env. Builds for Workers resolve it to the Workers environment (bindings included); everywhere else it is process.env:

import { env } from "agent-unit/env";

const storage = env.DB ? new D1Store({ binding: env.DB }) : new LibSQLStore({ url: env.DATABASE_URL });

They are merged into what agent-unit generates, never in its place: the names AGENT_UNIT_RUNS, AGENT_UNIT_INDEX, AgentUnitRun, AgentUnitIndex and the migration tag agent-unit-v1 are reserved, and the build says so if you reuse one. (nitro.cloudflare is the same setting and still works.)

In a Worker you write yourself:

import { createDurableAgentUnit } from "agent-unit/cloudflare";
import support from "./agents/support";

const unit = createDurableAgentUnit({ agents: { support } });

export const { AgentUnitRun, AgentUnitIndex } = unit;
export default { fetch: unit.fetch };
// wrangler.json: copy unit.wrangler (durable_objects bindings + migrations)

Inside an existing app

Already on Nitro (or something built on it)? Mount your agents with the module:

// nitro.config.ts
import { defineConfig } from "nitro";
import { agentUnit } from "agent-unit/nitro";

export default defineConfig({
  modules: [agentUnit({ basePath: "/api/agents" })],
});

In a Farm app, mount it with one catch-all API route, which works in farm dev and every production target (see examples/farm-app):

// src/app/api/ai/[...path]/route.ts
import { after } from "@farm.js/core/after";
import { agents } from "../../../../agents/server"; // createAgentUnit({ …, basePath: "/api/ai" })

const handle = (request: Request) =>
  agents.handler(request, { waitUntil: (work) => after(async () => void (await work)) });

export const GET = handle;
export const POST = handle;

Anywhere else that speaks Fetch (Hono, Next.js route handlers, Bun.serve, Deno.serve, a Worker):

import { createAgentUnit } from "agent-unit/server";
import { createStorage } from "unstorage";
import redis from "unstorage/drivers/redis";
import support from "./agents/support";

const unit = createAgentUnit({
  agents: { support },
  storage: createStorage({ driver: redis({ url: process.env.REDIS_URL }) }),
  basePath: "/api/agents",
});

export const fetch = (request: Request, ctx?: { waitUntil(p: Promise<unknown>): void }) =>
  unit.handler(request, { waitUntil: ctx?.waitUntil.bind(ctx) });

Pass the host's waitUntil where it has one. Without it, start and resume requests finish the work before they answer, which is the safe behaviour on hosts that freeze after a response.

Examples

Example	What it shows
`examples/universal`	The universal build: one `agents/` folder built for Node, Bun, Deno, Cloudflare, Vercel, Netlify and AWS Lambda, then every build run through the same pause, restart and resume flow (`npm run try`).
`examples/cloudflare-ai-sdk`	An AI SDK `ToolLoopAgent` (OpenAI and Claude) on Cloudflare Workers with the Durable Objects runtime: a refund pauses for approval, and `npm run verify` checks it against real models across a runtime restart.
`examples/farm-app`	Agents inside a Farm app at `/api/ai`, with an approvals inbox and live event stream at `/agents`, tested on the production server across a restart (`npm run smoke`).

Configuration

agent-unit.config.ts is optional:

import { defineConfig } from "agent-unit";

export default defineConfig({
  name: "support",                                   // manifest, MCP and A2A name
  storage: { driver: "redis", url: process.env.REDIS_URL },  // any unstorage driver
  preset: "vercel",                                  // or --preset
  runtime: "default",                                // or "durable-objects" on Cloudflare
  budget: "60s",                                     // yield and continue after this long
  sweep: "*/5 * * * *",                              // sweep schedule, or false
  basePath: "/api/agents",
  authorize: (request) => request.headers.get("authorization") === `Bearer ${process.env.API_TOKEN}`,
  retention: "30d",                                  // delete finished runs 30 days after they finish
  maxBodyBytes: 1_000_000,                           // largest accepted request body (default 1 MB)
  // agents: { support },                            // instead of the agents/ directory
  // adapters: [myAdapter],                          // extra framework adapters
  // nitro: { … },                                   // passed through to Nitro
});

Talk to your agents

HTTP and events

The full contract is in spec/. Events follow AG-UI (RUN_STARTED, TEXT_MESSAGE_CONTENT, TOOL_CALL_START, …) plus RUN_INTERRUPTED, RUN_SLEEPING and RUN_CANCELLED. Every event has a seq; it is the SSE event id, so EventSource reconnects with Last-Event-ID and picks up exactly where it left off.

Typed client

import { createAgentClient } from "agent-unit/client";

const agents = createAgentClient({ baseUrl: "https://agents.example.com" });

for await (const event of agents.run("refund", { orderId: "o_1" })) {
  if (event.type === "RUN_INTERRUPTED") console.log("needs approval:", event.interrupt.payload);
}
for await (const event of agents.resume(runId, { approved: true })) console.log(event.type);

await agents.list({ status: "interrupted" }); // an approvals inbox in one call
await agents.delete(runId);                   // remove a finished run and its history

The client reconnects dropped streams from the last event it saw.

MCP and A2A

POST /mcp speaks MCP over streamable HTTP: every agent is a tool, so Claude, Cursor or any MCP client can call your agents. A run that pauses returns its run id and the pending interrupt. GET /.well-known/agent.json is an A2A agent card with one skill per agent.

Testing

import { createTestUnit } from "agent-unit/testing";
import refund from "../agents/refund";

const unit = createTestUnit({ refund });
const parked = await unit.run("refund", { orderId: "o_1" });
expect(parked.status).toBe("interrupted");
const done = await unit.resume(parked.id, { approved: true });
expect(done.output).toBe("refunded");

Pass the same storage to a second createTestUnit to test a restart.

Customize

agent-unit is layered. Pick the layer that matches what you are changing:

You want to…	Layer	Who
Pause, sleep, keep state, make a call run once	Run primitives: `useRun()` / `run.step`, `run.interrupt`, `run.sleep`, `run.state`	App developers, in agent code
Choose host, storage, runtime, auth, time budget	`agent-unit.config.ts`	App developers
Mount the API in your own server, add routes, compose	`createAgentUnit()`, `createDurableAgentUnit()`, `createHandler()`	App developers
Support a new framework	`defineAdapter()` with `ctx.durable`, `ctx.kv`, `ctx.emit`	Framework and library authors
Run agents on a new kind of runtime	`RunEngine` options, or your own `RunService`	Runtime and platform authors

Mount and compose. createAgentUnit returns a Web handler, so it sits beside your own routes, middleware and auth:

import { createAgentUnit } from "agent-unit/server";

const unit = createAgentUnit({
  agents: { support },
  storage,                       // any unstorage instance
  basePath: "/api/agents",
  // Second argument: what the request does, with the run's agent and thread taken from storage.
  authorize: async (request, { action, threadId }) => {
    const session = await getSession(request);
    return session !== null && (threadId === undefined || threadId.startsWith(`${session.userId}:`));
  },
  budget: "60s",
});

// Any Fetch-style router, here Hono:
app.all("/api/agents/*", (c) => unit.handler(c.req.raw));

Build on the engine. Everything the HTTP API does goes through RunEngine, which you can drive directly (a queue consumer, a cron job, a test):

import { RunEngine, RunStore, resolveAgent } from "agent-unit/runtime";

const engine = new RunEngine({
  store: new RunStore(storage),
  agents: [resolveAgent("support", support, [myAdapter])],
  budgetMs: 60_000,
  // Hand wake-ups to your own scheduler (a delayed queue message, a job runner…) instead of timers:
  scheduleWake: (runId, at) => scheduler.at(at, () => engine.handleWake(runId)),
});

const { run, done } = await engine.start("support", { prompt: "refund o_1" });

A new runtime. createHandler(service) serves the whole API (runs, SSE, MCP, A2A) from any object that implements RunService from agent-unit/server. The Durable Objects runtime is one such implementation: it routes each call to the run's object instead of running it in-process.

Adapters for other frameworks

AI SDK, Mastra, OpenAI Agents and LangGraph are supported out of the box. Any other framework can publish its own adapter package and maintain it on its own schedule: no change to agent-unit, no pull request. To use one:

npx agent-unit add agent-unit-adapter-crew

It installs the package, checks that it is an adapter, and adds two lines to agent-unit.config.ts (asking first; --yes skips the question). Adding them by hand works the same:

import crewAdapter from "agent-unit-adapter-crew";
export default defineConfig({ adapters: [crewAdapter()] });

Adapters in the config are tried before the built-in ones, so a framework's own adapter replaces agent-unit's.

Write and publish an adapter

An adapter teaches agent-unit one framework. Most are a few dozen lines:

// agent-unit-adapter-crew/src/index.ts
import { defineAdapter } from "agent-unit/adapter";
import { CrewAgent } from "crew";

export default function crewAdapter() {
  return defineAdapter<CrewAgent>({
    name: "crew",
    apiVersion: 1, // the adapter interface this was written for
    match: (value): value is CrewAgent => value instanceof CrewAgent,
    describe: (agent) => ({ description: agent.description, tools: agent.tools.map(({ name }) => ({ name })) }),
    async run(agent, ctx) {
      const model = ctx.durable.model(agent.model); // journaled, streams text events
      const tools = ctx.durable.tools(agent.tools); // journaled, emits tool events
      return agent.run(ctx.input, { model, tools, signal: ctx.signal });
    },
  });
}

To publish it:

Export a function returning the adapter as the default export, so agent-unit add can find it.
Name it agent-unit-adapter-<framework>, or @your-scope/agent-unit, and add the agent-unit-adapter keyword, so people can find it on npm.
Declare agent-unit and your framework as peer dependencies, so the app's copies are used.
Check it with the same checks agent-unit runs on its own adapters, in any test runner. Report every real side effect through kit.effect; the checks fail when one runs outside a journaled call, or runs again after a restart:

import { assertAdapter } from "agent-unit/testing";

test("agent-unit adapter", () =>
  assertAdapter({
    adapter: crewAdapter(),
    agent: (kit) => new CrewAgent({ model, tools: [refundTool(() => kit.effect("refund"))] }),
    // Optional: an agent that pauses through your framework, and the answer that resumes it.
    pause: { agent: (kit) => approvalAgent(kit), answer: { approved: true } },
  }));

See spec/adapters.md for adapters of frameworks that keep their own state.

Several servers, one store

Run as many instances as you like over one store. With Redis it is exact out of the box:

storage: { driver: "redis", url: process.env.REDIS_URL },

With createAgentUnit on your own server, wrap the storage yourself:

import { RunStore } from "agent-unit/runtime";
import { redisCoordination } from "agent-unit/redis";
import { createStorage } from "unstorage";
import redis from "unstorage/drivers/redis";

const driver = redis({ url: process.env.REDIS_URL });
const storage = new RunStore(createStorage({ driver }), redisCoordination({ client: () => driver.getInstance!() }));
// createAgentUnit({ agents, storage })

redisCoordination needs one method, eval, so any client with Lua scripting works (ioredis, node-redis behind a small wrapper). For another store, implement AtomicWrites.compareAndSet (save a run record only if its version is still the one read) and LeaseBackend with that store's own conditional write, such as a database UPDATE … WHERE version = ?.

Limits

Leases make one execution run a run at a time, renewed while it works, and every change to a run record is a versioned save: it lands only if nobody else saved the run since it was read, so two processes racing to resume, cancel or wake a run cannot both win (the loser gets a 409). On Redis both are atomic, wired in automatically by agent-unit build and agent-unit dev when storage uses the redis driver. On other shared stores the save is check-then-write, which is exact within a process and best effort across processes racing within milliseconds; pass atomic leases and atomic writes to RunStore for those (see below), or use the Durable Objects runtime, where each run's object is its only executor and saves go through the index atomically.
A Mastra agent with memory reads the run's thread (and input.resourceId, or the thread, as its resource) and stores each finished turn once. With working or observational memory turned on, Mastra saves messages itself as the turn runs, so a run that pauses or recovers from a crash may store that turn twice.
In the Durable Objects runtime, the run list and shared state live in one index object, written when a run starts, parks or finishes (not on every step). agent-unit dev runs the default runtime locally.
Streamed text and tool-argument deltas that arrive within 50ms of each other are merged into one event, so a long answer costs a few writes instead of one per token, and live clients see text up to 50ms later. deltaBatchMs: 0 on createAgentUnit keeps every delta its own event.