MCP server de Lousho

Lousho — TypeScript AI agent SDK

TypeScript AI agents you can test in CI, pause for approval, and resume after a crash. Runs in your own Node process.

Why Lousho

• Test agents in CI without an API key. lousho eval records a real run once, replays it with no network, and reports when the tool-call trajectory drifts. Evals
• Resume after a crash. Give the agent a store and agent.resume() finishes an interrupted run. Tool calls that already have a result do not run again. Durable execution
• Pause for a human. A tool marked needsApproval pauses the run. Another request, or with a durable store another process, can approve it later, and the run continues where it stopped. Approvals
• Runs in your own Node process. There is no hosted runtime. lousho build targets a Node server, Docker or Cloudflare Workers, and lousho studio runs on your machine. Deployment
• Works with the Vercel AI SDK. fromAiSdk() and fromAiSdkTools() wrap AI SDK models and tools, so you can move over one piece at a time. Migrating from the AI SDK

Try it in 60 seconds, no API key

You need Node.js 22.19 or newer and a network connection to install packages. You do not need an API key.

npm create lousho@latest my-agent -- --yes
cd my-agent
npm test

--yes picks the defaults: the minimal template and the OpenAI provider. npm test runs the project's tests on a mock model, offline.

Now add an eval. Save this as src/weather.eval.ts:

import { createAgent, defineEval, defineTool } from 'lousho';
import { mockModel } from 'lousho/testing';
import { z } from 'zod';

const getWeather = defineTool({
  name: 'get_weather',
  description: 'Current weather for a city',
  input: z.object({ city: z.string() }),
  execute: async ({ city }) => ({ city, tempC: 21, sky: 'sunny' }),
});

defineEval({
  name: 'weather advice',
  agent: () =>
    createAgent({
      instructions: 'Check the weather before giving advice.',
      tools: [getWeather],
      provider: mockModel([
        { toolCalls: [{ name: 'get_weather', args: { city: 'Lisbon' } }] },
        { text: 'It is sunny and 21°C. Wear a T-shirt.' },
      ]),
    }),
  async test(t) {
    await t.send('What should I wear in Lisbon today?');
    t.completed();
    t.calledTool('get_weather', { args: { city: 'Lisbon' } });
    t.maxSteps(4);
  },
});

Then run it:

npx lousho eval

EVAL            CASE  RESULT  SCORES  DURATION
weather advice  -     PASS    -       14ms

1 eval(s): 1 passed, 0 failed, 0 with soft failures (14ms)

To test against a real model, use model: 'openai/gpt-4.1-mini' in place of provider, set your key and record once. This step needs the key. Commit the cassettes it writes. CI then replays them with no network and no key:

npx lousho eval --record
npx lousho eval --replay
npx lousho eval --drift

--drift re-records each case and shows where the tool calls, arguments or step count changed. See Record, replay and drift.

How it compares

	Lousho
Where it runs	In your own Node process. No hosted runtime.
Tests without an API key	mockModel scripts, and recorded cassettes that lousho eval --replay plays back offline
Crash recovery	agent.resume() on memory, file, SQLite or Cloudflare KV stores
Human approval	needsApproval pauses the run; with a durable store, another process can resume it
License	MIT

For a side-by-side look at the Vercel AI SDK, OpenAI Agents SDK, Mastra and LangGraph.js, see the comparison page.

Installation

Requires Node.js 22.19 or newer. To start a new project, see Try it above. To chat with that agent, copy .env.example to .env, put your API key in it and run npm run dev.

To add Lousho to an existing project:

npm install lousho ai@^7.0.0 zod @ai-sdk/openai@^4.0.0

See Installation for other providers. npx lousho doctor checks your setup and prints a fix for anything missing.

Quickstart

This file uses top-level await, so it must be an ES module (.mts, or "type": "module" in package.json). Set OPENAI_API_KEY to run it.

import { createAgent, defineTool } from 'lousho';
import { z } from 'zod';

const getWeather = defineTool({
  name: 'get_weather',
  description: 'Current weather for a city',
  input: z.object({ city: z.string() }),
  execute: async ({ city }) => ({ city, tempC: 21, sky: 'sunny' }),
});

const agent = createAgent({
  model: 'openai/gpt-4.1-mini',
  instructions: 'You are a travel assistant. Check the weather before giving advice.',
  tools: [getWeather],
});

const { text } = await agent.send('What should I wear in Lisbon today?');
console.log(text);

Agent folders

An agent can also be a folder, which you run with npx lousho dev ./my-agent:

my-agent/
├── agent.json          # Optional: model, permissions, limits
├── instructions.md     # The system prompt
├── tools/              # defineTool() tools
└── skills/             # Instructions loaded on demand

Loading a folder runs its code, so only load ones you trust. See Agent directories.

Features

• MCP: connect to MCP servers, or serve your agent as one. MCP
• Multi-agent: sub-agents that run in parallel, and handoffs to specialists.
• UI: bindings for React, Vue and Svelte, plus AI SDK useChat and Next.js routes.
• Tracing: OpenTelemetry GenAI spans and a local trace viewer. Observability
• Agent Forge: lousho studio opens a local visual canvas, run debugger and chat. Agent Forge
• Self-healing (experimental): lousho improve turns an agent's failed runs into tested edits to its instructions and skills, applies one that passes (or queues it for a person), and rolls it back on a regression. Self-healing
• Workspaces (experimental): lousho/workspaces holds the contracts for a workspace of pages, harnesses and shared policy: schemas, validation, harness composition and typed authoring operations. Workspaces
• Coding agents: workspace tools, permission rules and starter kits from lousho add. Build a coding agent

Documentation

The full guides are at lousho.com, in English and Arabic. Good places to start: Quick Start, Tools, Providers, CLI and Troubleshooting.

Coding agents can read the docs from node_modules/lousho/docs, indexed by llms.txt. The repo also has every page in one file, llms-full.txt.

Examples

Most examples run offline with a mock provider. The examples index has the full list and how to run each one. A few to start with:

• ops-pipeline: monitor alert, Slack button, human approval, fixer agent, gated GitHub PR
• coding-harness: a coding agent with checkpoints, permission rules and a loop guard
• deep-research: parallel researcher sub-agents and a cited report
• support-desk: triage handoffs and an approval-gated refund tool

Status

Alpha. The API may change before 1.0. Pin a version. Upgrading covers the breaking changes between alphas and how to move your code.

Community

• Questions and ideas: GitHub Discussions.
• Bugs and feature requests: GitHub issues.
• Arabic docs: lousho.com/ar.

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup and the checks a pull request must pass. To report a vulnerability, see SECURITY.md.

License

MIT © Ali Mohammad