Clatter SDK

Introduction

Clatter SDK is the official native toolkit for the ClatterAI platform. It runs entirely on your machine — no account required to get started — and gives developers a complete workshop for designing, testing and shipping AI agents. Typical use cases include customer-support agents, internal copilots, automated workflows and reasoning pipelines. Clatter SDK is proprietary software, free for local development.

Installation

Install the Clatter CLI on your machine — one line, no dependencies.

macOS & Linux

curl -fsSL https://sdk.clatterai.com/install.sh | sh

Windows (PowerShell)

iex (irm https://sdk.clatterai.com/install.ps1)

After installation, verify it works:

clatter --help

What's in the kit

Unlike typical SDKs that ship only a client library, Clatter installs a complete local toolkit:

• clatter — the CLI: scaffold, run, simulate, deploy.
• @clatter/sdk — the TypeScript / Python SDK to define agents in code.
• Agent Studio — a local web UI (clatter studio) to visually compose tools, prompts and memory.
• Prompt Simulator — replays thousands of past conversations against a new prompt in seconds.
• Time-travel Debugger — step backward through any agent trace and replay forward.
• Local runtime — a snapshot-based agent runner with sub-50ms cold starts.

Quick Start

Create a project, define your first agent, and run it locally.

clatter init my-agent
cd my-agent
clatter dev

Open the Studio in your browser:

clatter studio    # opens http://localhost:7878

Configuration

Every project has a clatter.config.json file:

{
  "name": "support-bot",
  "runtime": "local",
  "region": "auto",
  "model": "gpt-5",
  "memory": { "backend": "local-kv", "ttl": 86400 },
  "secrets": ["OPENAI_API_KEY", "DATABASE_URL"]
}

Agent Studio

Studio is a visual canvas that runs locally on localhost:7878. Drag tools, prompts and memory blocks into a live graph. Edits hot-reload into the running agent without restarting the process.

clatter studio --port 7878

Prompt Simulator

Replay historical traces against a candidate prompt to catch regressions before deploy.

clatter simulate ./traces.jsonl --prompt v2 --diff

Time-travel Debugger

Step backward through any agent run, mutate intermediate state, and replay forward to see the new outcome — a first in agent tooling.

clatter debug run_2f8a91 --step -3 --mutate "tools.search=stub"

Deploying Agents

Push to the Clatter runtime. Each deploy is atomic and instantly rollbackable.

clatter deploy --env production
clatter rollback --to v12

CLI Reference

SDK API

Node.js / TypeScript:

import { defineAgent } from "@clatter/sdk";

export default defineAgent({
  name: "support-bot",
  model: "gpt-5",
  tools: ["search", "tickets"],
  handler: async ({ message, tools }) => {
    const ctx = await tools.search(message);
    return tools.reply(ctx);
  },
});

Python:

from clatter import Agent

agent = Agent(
    name="support-bot",
    model="gpt-5",
    tools=["search", "tickets"],
)

@agent.handler
async def handle(message, tools):
    ctx = await tools.search(message)
    return await tools.reply(ctx)

Performance & Scaling

Clatter uses snapshot-based cold starts, adaptive batching and a shared inference pool to deliver p50 cold starts under 50ms and roughly 10× lower compute usage than traditional container deployments. Local development uses the same runtime — what you measure on your laptop is what you get in production.

FAQ