Cloudwright

August 1, 2026 · View on GitHub

Describe a cloud architecture in English. Get Terraform, costs, and a compliance check.

PyPI CI License: MIT Python 3.12+ xmpuspus/cloudwright MCP server

A terminal runs cloudwright init, cost, compliance and plan. The compliance table maps every finding to a HIPAA and SOC 2 control ID. The plan step ends on a DEPLOYABLE verdict from terraform validate.

pip install 'cloudwright-ai[cli]'
export ANTHROPIC_API_KEY=sk-ant-...
cloudwright design "HIPAA healthcare API on AWS with Postgres and Redis"

Cloudwright turns one line of English into a typed spec, a cost breakdown, a control-mapped compliance report, and infrastructure code. It covers AWS, GCP, Azure and Databricks across 114 service keys. Only design, modify, chat and adr call a model. Every other command runs offline and needs no API key.

Quickstart · Compliance · Agents · Docs · Changelog

A prompt produces a spec, a cost, a control-mapped report, and Terraform

  • Spec. Typed YAML you commit, diff and review. Everything below reads from it.
  • Cost. Per component and region-aware, with a confidence flag on every line.
  • Compliance. HIPAA, SOC 2, PCI-DSS, FedRAMP, GDPR, ISO 27001 and NIST 800-53 control IDs.
  • Infrastructure code. Terraform, OpenTofu, Pulumi (TypeScript or Python) and CloudFormation.
  • Diagrams. ASCII, Mermaid, D2, and a web canvas you can edit by hand.
  • An MCP server, so any coding agent runs the same checks inside its own loop.

Exports carry safe defaults. S3 gets a public-access block, SSE and versioning. RDS gets encryption, multi-AZ and deletion protection. EC2 gets IMDSv2. A compliance framework overrides the workload profile, and always forces encryption and high availability.

Quickstart

cloudwright design "HIPAA healthcare API on AWS with Postgres and Redis"
cloudwright cost spec.yaml --workload-profile medium
cloudwright compliance spec.yaml --frameworks hipaa,soc2
cloudwright export spec.yaml --format terraform -o ./infra
cloudwright plan spec.yaml --target terraform          # proves it deploys, never applies
cloudwright chat --web                                 # canvas at http://localhost:8765

Add --json before any subcommand for machine-readable output, or --stream to watch tokens arrive. Set ANTHROPIC_API_KEY or OPENAI_API_KEY for the four commands that need a model.

Every finding carries the control ID it violates

The web canvas Compliance tab shows a per-framework table for HIPAA, SOC 2 and FedRAMP. Each row carries satisfied and violated control IDs. The Plan tab then returns a DEPLOYABLE verdict.

Other tools scan infrastructure after you deploy it. Cloudwright maps each finding to its control before any resource exists. The fix then costs a spec edit, not a change ticket. HIPAA 164.312(a)(2)(iv), SOC 2 CC6.1 and FedRAMP SC-28 come from the built-in scanner, with no extra tooling. Checkov folds into the same report when it sits on your PATH.

  • --oscal writes an OSCAL 1.1.2 component-definition with deterministic UUIDs.
  • --traceability prints the chain from component to resource to control to status.
  • cloudwright plan runs terraform validate against the export, and never applies.

The review needs no API key and no network

cloudwright review prints a severity-ranked table for the patient-portal spec. It scores the spec 39 out of 100, with 15 findings and 8 blocking. A second command traces each component and Terraform resource to a violated HIPAA control.

cloudwright review runs the scorer, the linter and the validator over a spec, and returns one severity-ranked report. The same three critics run inside cloudwright design. When blocking findings survive generation, the architect repairs the spec once and records the change in spec.metadata.critique. Pass Architect(repair=False) to turn that off.

Canvas edits never call the model, so they are instant and free

The web canvas with a boundary-aware diagram. A catalog drawer adds an ElastiCache node, a side panel edits its label and config, and the cost total updates.

Add, drag, connect, edit and delete are deterministic frontend mutations. The Catalog drawer serves the resource list per provider and five approved multi-resource modules. Its standards check flags orphan connections, partial modules and missing tags. An intact module exports as a single Terraform module block, with the catalog's pinned source and version.

One MCP server reaches 11 coding agents

cloudwright integrate --harness claude-code       # exact wiring, in that client's format
cloudwright integrate --harness cursor --write    # merge it into the right file
cloudwright integrate --rules --agent-file claude # a gate block for CLAUDE.md

Do not hand-write the config. cloudwright integrate emits it for Claude Code, Cursor, Cline, Windsurf, GitHub Copilot, Zed, Codex CLI, Junie, Kiro and Antigravity. Aider gets a CLI-pipe recipe instead, because it speaks no MCP. Every client wants a different shape: Zed wants context_servers, Copilot wants servers, and Codex wants a TOML table.

The server exposes 22 tools in 9 groups: design, cost, validate, analyze, export, session, review, compliance and plan. Full matrix in docs/integrations.md.

Nine offline commands grade, scan and compare a spec

lint runs 10 anti-pattern checks. score grades 5 dimensions. analyze reports blast radius and single points of failure. policy enforces policy-as-code with 9 built-in rules. security scans the spec and the exported HCL. drift compares a design against a tfstate, and --remediate turns the gap into a cost, compliance and plan preview.

review, compliance and plan are above. See docs/cli-reference.md.

Python API

from cloudwright import ArchSpec
from cloudwright.cost import CostEngine
from cloudwright.validator import Validator
from cloudwright.exporter import export_spec

spec = ArchSpec.from_file("spec.yaml")
priced = CostEngine().estimate(spec, workload_profile="medium")
findings = Validator().validate(spec, compliance=["hipaa", "pci-dss"])
hcl = export_spec(spec, "terraform", output_dir="./infra")

v1.9.0 gave the diagram arrowheads and got the canvas off React Flow's defaults

Measured by driving all 16 init templates through the running app, and 8 interactions at 1920, 1440 and 390px.

  • Every connection draws an arrowhead. The computed markerEnd was none on every edge.
  • Connection lines clear 3:1 contrast. They ran at 1.42:1 in light and 1.81:1 in dark.
  • The canvas fits a phone. The 0.5 zoom floor left 2 of 8 nodes off the pane at 390px; the floor is 0.12.
  • A drag inside a VPC pans the canvas. It used to drag the box out of shape and save nothing.
  • A node moves the distance you drag it. A tight boundary plus extent: "parent" allowed 5 to 10px.
  • Connections stop hiding behind cards. 17 of them did across the 16 templates; 5 still do.

Earlier releases added control-ID mapping and plan (v1.5.0), the self-correcting architect and OSCAL (v1.6.0), cloudwright integrate (v1.7.0), and the dark theme and responsive layout (v1.8.0). Full history in CHANGELOG.md.

Compatibility

  • Python 3.12+
  • Models: Anthropic (Claude Sonnet, Haiku) and OpenAI (GPT-5+ family), auto-detected from env.
  • Clouds: AWS, GCP, Azure, Databricks. 114 service keys total.
  • Install variants: cloudwright-ai[cli], cloudwright-ai[web], cloudwright-ai-mcp.

Contributing, license, changelog