Requirements in, architecture out

Describe your app. Get the architecture.

Type your requirements or upload a document. An agent designs the services, how they connect and how changes ship, on AWS, Azure, Google Cloud or open source.

Attach Region Planner

Word, PDF, Excel, Markdown or text, up to 10 MB · Ctrl+Enter to start

Try a sample

Detail A · How it works

From brief to drawing

Read the brief

Type it, or drop in a Word, PDF or Excel file. Paragraphs and tables are read alike, and nothing leaves your browser.

Mark up what matters

The numbers that size a design are picked out: users, traffic, data, uptime, region, compliance and budget. Anything missing becomes a question for you.

Draw the architecture

An agent chooses each service with MCP tools, from the edge to the data, and says why, quoting your own words.

Add build and deploy

Source control, CI, a registry, a release pipeline and infrastructure as code come with every design.

Trace the workflows

A request, a release, a background job: each one is a numbered path through the drawing.

Every cloud, with costs

The same design is drawn for AWS, Azure, Google Cloud and open source, priced from one month to three years.

Read
Understand
Design with MCP tools
Add build and deploy
Describe workflows
Map to every cloud

Sheet 04 · How to

How to use Clarchy

From a brief to a design you can review, price and share, in six steps. Each one matches what you see on screen.

1Write the brief

Describe the app in a few plain sentences, or attach a Word, PDF, Excel, Markdown or text file of up to 10 MB. Tables are read as well as paragraphs. Each fact below changes the design; anything you leave out becomes an assumption or a question you can answer later.

Example brief

An assistant that answers employees' questions from our HR policy documents, for about 8,000 employees in Germany. Employees sign in with SSO. It needs 99.9% uptime, GDPR applies, and chat history must be kept for 2 years. The budget is $4,000 a month, and our 6 developers already use GitHub Copilot.

Clarchy picks out 8,000 users · Frankfurt region · 99.9% availability · GDPR · records kept 2 years · $4,000 a month · 6 developers

Say	For example	What it changes
How many people use it	about 8,000 employees 120,000 patients	Sizes every service. Below 50,000 users the API runs as serverless functions, above that on containers. Without a number, Clarchy plans for 10,000.
The busiest moments	300 requests per second 2,000 orders an hour	Sizes the API and the database for the peak.
How much data	2 TB of documents grows 50 GB a month	Sizes storage and its cost.
Uptime	99.95% uptime 24/7	At 99.9% or more, services and databases are spread across zones. Without a number, Clarchy assumes 99.9%.
Where users are	users in Germany most customers are in India	Picks the region on every cloud, and the data-protection laws that apply, such as GDPR or India's DPDP Act.
Rules that apply	HIPAA applies PCI DSS SOX	Adds audit logging, key management and access control, and checks each rule on the Policies tab. Clarchy knows HIPAA, PCI DSS, GDPR, SOX, SOC 2, ISO 27001 and FedRAMP.
How long to keep records	keep records for 7 years	Adds backups, and archive storage for the retention period.
Budget	$4,000 a month $50,000 a year	Saved with the design, so you can hold the estimate against it.
What you already run	the backend is already in Docker we run Kubernetes	Keeps it: Docker means containers, Kubernetes means Kubernetes.
Your team and tools	6 developers GitHub Copilot Codespaces	Adds the developer tools you name, priced per developer.
AI features	an assistant that answers from our policy documents LangGraph agents Langfuse LiteLLM	Adds a model, document search, agent orchestration, an LLM gateway, tracing and guardrails, as the brief needs.

2Plan it

On Plan, type or paste the brief, or click Attach. You can also drop a file on the box.
Region: leave From the text to use the place your brief names, or choose one.
Planner: the rule-based planner is instant and keeps everything in your browser. To have an AI model write the design, see Use an AI model.
Click Design architecture, or press Ctrl+Enter. The first plan downloads the planning engine, about 15 MB; after that, plans take seconds.
Six steps run: read, understand, design, add build and deploy, describe the workflows and map to every cloud. Show steps lists what was found at each one.

TipNo brief yet? Try a sample, under the box, runs a ready-made one.

3Read the drawing

Pick a cloud with the tabs above the drawing: AWS, Azure, Google Cloud or Open source. It is the same design, drawn in each cloud's own style.
Click any service to see why it is there: the sentence of your brief that asked for it, how it connects, its sizing, its monthly cost and the alternatives.
Actual size shows the drawing at full scale. Esc closes the details.
The title block under the drawing lists the cloud, the region and the number of services.

4Check cost, policies and workflows

The tabs next to the cloud names show the rest of the design.

Services: Every part of the design with the service it uses on this cloud, grouped by stage, and why it is there.
Cost: The monthly estimate, and totals for 1 month, 6 months, 1 year and 3 years, on demand and with commitments. Add startup or model credits to see what is left to pay. AWS prices come from the AWS Price List API; Azure, Google Cloud and third-party prices are marked approximate.
Policies: The AI and data rules that apply, such as the EU AI Act, GDPR or HIPAA. Each obligation is covered by a component, a gap with a suggestion, or an action for your team. A checklist for design reviews, not legal advice.
Workflows: Numbered paths through the drawing: a user's request, a release, a background job. Press Play to walk through one step by step.
Spec: The YAML behind the drawing. Edit it and the drawing, services and costs update as you type; Reset brings back the plan.

5Answer questions and re-plan

The bar above the drawing lists the questions your brief left open and the assumptions Clarchy made instead.
Click Answer and re-plan. Answer in your own words, such as 50k, yes or Frankfurt, or correct an assumption.
Clarchy plans again with your answers. Answered questions are not asked again, and each round keeps the answers you gave before.

AUse an AI model

In Planner, choose Open-source model (Hugging Face).
Paste a Hugging Face token that may make calls to Inference Providers (create one free ↗), and keep or change the model.
For a model on your own machine, open Use another endpoint: run Ollama with OLLAMA_ORIGINS=*, enter http://localhost:11434/v1 and a model such as qwen2.5:7b, and leave the token empty.

PrivacyThe token goes from your browser straight to the model provider and is forgotten when you close the tab. With the rule-based planner, nothing leaves your browser.

BCompare clouds

Services has a card for each job a system needs, with the service that does it on every cloud. Filter by stage or cloud; Close and Partial mark differences worth knowing.
Examples are reviewed reference architectures. Open one to see it on every cloud, with costs.
The Blog walks through worked examples, such as a work-item assistant for 10,000 employees.

Start a plan