An app whose agents each have a computer, an address and a budget.
Praecise Harness writes the app as a folder. Praecise Cloud runs it, gives each agent a Silo to work in, an Inbox address to be reached at and Skerry limits to pay inside, and Engine accelerates the models served on Praecise Metal.
In short
AI Apps on Praecise Cloud are written with Harness, run with a Silo, an Inbox address and Skerry limits for each agent, and served with Engine on Metal. This page is for people building agent apps. Start with npx @praecise/harness init my-app.
00
At a glance.
Open source
Harness and Engine are public.
Both are Apache-2.0 repositories under github.com/Praecise.
In use
Praecise runs on them.
Harness runs Praecise's agents and Engine accelerates the models served on Praecise Metal.
Live
Sites on the same platform.
pavilionbygold.com and tenzro.com are deployed on Praecise Cloud.
01
What an AI app gets here.
- 01
Write the agent as a file
The agent is reviewable like any other code.
An agent is a TypeScript file in agents/ with a role and the tools it may use. Diff it, review it and version it with the rest of the app.
- 02
Ground it in your own documents
Answers come from files you control.
Put .md and .txt files in memory/ and every agent can answer from them. Ingest PDF, Word, Excel, PowerPoint, CSV and images into a store.
- 03
Pick quality, not a model
One dial instead of a model id to keep current.
Set quality to fast, balanced or best. Each question goes to the cheapest model that can answer it, and a request may be capped with a ceiling.
- 04
Stop for a person
Risky steps wait for approval.
A workflow can pause at an approve step, and resume where it stopped. Approvals are attributed, and signed when you wire a signer.
- 05
Give it a computer
Code runs away from your servers.
An agent that must run commands gets a Silo: its own isolated computer with Python, Node, Git and a compiler ready, and a receipt for each run.
- 06
Give it an address
People and agents can write to it.
Praecise Inbox gives an agent a mailbox on your verified domain, with everything said kept in one record a person can search.
- 07
Give it a budget
An agent pays only inside limits you set.
Skerry keys are scoped per agent, with spend limits, approvals and revocation. Agent payments are sandbox today, with live keys on approval.
- 08
Serve the model
Faster generation on structured output.
Praecise Engine sits between the app and a backend and decides whether speculation pays. Measured on structured text only: see the Engine README.
02
Start with a folder.
Start a new app, install it and run it locally. The dev server serves a dashboard, a chat page for each agent, a REST API, an MCP endpoint and an A2A peer.
npx @praecise/harness init my-app
cd my-app
npm install
npx @praecise/harness dev
my-app/
agents/support.ts one file per agent
functions/refund.ts your code, callable by an agent
memory/faq.md what every agent can answer from
guard.ts which tool calls the app actually makes
.env PRAECISE_API_KEY=...
// agents/support.ts
import { agent } from "@praecise/harness";
export default agent({
role: "Customer support for Acme.",
tools: ["refund"],
});Node 22 or newer. With one key in .env, the app runs on the models and stores Praecise Cloud provides. To use your own endpoints, describe them in praecise.config.ts.
03
Build, Place, Operate.
Build
Write the folder.
Agents, workflows, memory, functions and stores live in one folder. Run praecise doctor to see everything wrong with it in one pass.
Place
Put it behind the edge.
A Harness app is a Node app. It is placed with a domain and certificate, and its environment settings are kept outside the release.
Operate
Watch releases and runs.
Releases are health-checked and reversible, and the dev server's traces and the run record show what each request did.
04
What people build.
- 01
A support agent with its own inbox
It answers from your documents, replies from its own address and escalates anything about a chargeback to a person.
- 02
A booking agent that takes payment
It takes a booking and settles through Skerry, in sandbox today, inside a spend limit and with approvals.
- 03
A document intake workflow
It reads PDFs and spreadsheets into a store, extracts fields and waits for a person to approve before it acts.
- 04
A coding agent in a Silo
It runs commands and one program against bound services, with no network and a receipt for every run.
- 05
A model served on Metal
Engine speeds up an open model for structured work such as SQL and code, on hardware Praecise operates.
- 06
An MCP service for other agents
Every Harness app is also an MCP endpoint and an A2A peer, so another agent can call it.
05
Next pages.
06
Praecise accounts are by invitation. Open the console, or talk to us.
New to Praecise Cloud? Accounts are set up with the Praecise team: talk to us, then sign in to the Cloud section of Praecise Spaces with an email code.