An API for agents with things to do

Big ideas.
Small goat.
Real computers.

Managoat is open-source infrastructure for running agents through an API. Give Claude Code, Codex or Gemini CLI a sandbox, files and tools. Build whatever delightful thing you had in mind on top.

Self-hosted. Apache-2.0. Mildly goat-themed.

Field notes / 001 The useful bit underneath

Your wonderfully specific app

A button. A chat. A cron job. Your call.

Managoat

Starts the machine. Runs the agent.
Keeps track of the conversation.

One conversation’s sandbox

A coding agent + a real filesystem + a shell

Your repositories, packages, skills and tools.

Goat for scale. Computer for work.

A few things we’ve built

The API is plumbing.
The possibilities are the point.

Here are a few things we’ve built on it. Real apps, open source, with different ideas about what an agent ought to feel like. Borrow liberally.

01

Briefing Room

Describe what you need to understand and why. A research agent reads sources on the web and returns a cited brief instead of a chat transcript.

Under the hood: web research that returns a document instead of a chat interface.

See how it’s built →
02

Table Talk

Drop in a CSV. An analyst runs Python in its sandbox and returns charts with plain-language findings. Keep asking questions of the same data.

Under the hood: sandboxed Python analysis behind a file-upload interface.

See how it’s built →
03

Repo Sage

Name any public GitHub repository. An agent clones it in a sandbox and answers with file-and-line citations that link back to the source.

Under the hood: repository checkout, search and file-and-line citations.

See how it’s built →
04

Mission Control

Describe a goal. A coordinator drafts a plan for your approval, then starts one sandboxed agent per task. Watch them work and receive one combined report.

Under the hood: approval gates, parallel runs, combined event streams and a final report.

See how it’s built →
Visit the whole menagerie →

Exhibit A

A workshop
for ambitious work.

Workbench turns projects, tasks and agents into a place to get things done. Assign a task, follow the work, and pick up the conversation.

Its interface is an app built on Managoat’s public API. The same building blocks are there for your own workshop, research desk, or next big idea.

Peek inside Workbench →
Workbench with a storefront work item and Builder assigned to it
Workbench: projects, tasks and agents in one place.

Yes, there is an API under the goat

Send a task.
Get some actual work back.

Your app sends the prompt. Managoat provisions a sandbox and runs the agent. Stream its progress, read the result, and send a follow-up to the conversation.

Configure reusable agents and environments, connect MCP tools, and provide credentials for the job. Managoat handles the machines and their lifecycle.

hello-goat.ts@managoat/sdk
import { Managoat } from "@managoat/sdk";

const managoat = new Managoat();
const run = await managoat.runRequest({
  agent_id: "your-agent-id",
  prompt: "Explore this repo. Draw me a map.",
});

console.log(run.text);

After creating an agent, set MANAGOAT_BASE_URL and MANAGOAT_API_KEY in your server’s environment. Replace the example agent ID with yours.

Go from idea to running agent

A fresh sandbox.
In seconds.

Send a prompt and Managoat gets a computer ready for the job. Measured median sandbox provisioning and setup: 4.7–5.0 seconds for Claude Code and 9.2–9.8 seconds for Codex, across Sprites and E2B.

Then the runtime starts and the agent gets to work. The first-text column includes that whole path from the prompt arriving at Managoat.

Median cold-start timings by runtime and sandbox provider
Runtime / sandbox Sandbox + setup Prompt → first text
Claude Code / Sprites 4.7s 8.6s
Claude Code / E2B 5.0s 8.0s
Codex / Sprites 9.8s 20.7s
Codex / E2B 9.2s 13.9s

Measured medians · 15–16 October 2026. Timings vary by model and workload.

Some assembly encouraged

Bring your own bright idea.

Run Managoat on your infrastructure, bring your model credentials, and make it yours.