Strata MCP Server
Enables LLMs to model, validate, and analyze multi-cloud infrastructure as a typed graph, with tools for IaC import/export, cost estimation, and architecture review.
README
Strata
Live app: strata.mk-dir.com

A registry-driven canvas for modeling cloud infrastructure as a typed graph.
Drag services onto a canvas, configure them with auto-generated forms, and connect
them with typed relationships (contains, routes_to, invokes, peers_with, …)
— with a multi-cloud catalog spanning AWS, Google Cloud and Azure. It can also import
existing Infrastructure-as-Code — CloudFormation (JSON/YAML), Terraform
show -json (AWS/GCP/Azure) and Azure ARM — into the same graph, export the graph
back out as IaC, and ingest live cloud state via provider discovery APIs.
On top of the graph sits a layer of pure analysis engines — validation with one-click autofix, true internet-reachability evaluation, an "Explain & Clean" account review (cost + risk + a safe-cleanup checklist), cross-cloud migration mapping, a change/audit receipt between any two versions, tag tinting/filtering, canvas annotations, and a Diagram-as-Code DSL — all exposed both in the UI and over MCP.
MCP server
An LLM/agent can drive Strata over the Model Context Protocol.
The server (src/mcp/server.ts) is a thin wrapper over the same pure engines the
app uses — registry, validation, IaC import/export, and cost — so an agent can
query and transform a graph exactly as the UI does. No DOM, network, or
credentials.
Run it (stdio transport; uses npx tsx, no build step):
npm run mcp
Then point an MCP client at that command, e.g.:
{
"mcpServers": {
"strata": { "command": "npm", "args": ["run", "mcp"], "cwd": "/path/to/strata" },
},
}
Tools: list_services, get_service, validate_architecture,
suggest_rules, import_iac, export_iac, estimate_cost,
review_account (cost + risk + safe-cleanup report), evaluate_reachability
(true internet-reachability, not topology), map_to_cloud (cross-cloud
equivalence), graph_to_dsl / graph_from_dsl (Diagram-as-Code round-trip),
list_autofixes / apply_autofix (deterministic validation fixes),
change_receipt (audit receipt of what changed between two graphs),
tag_report.
Live cloud discovery is separate — it runs through a Cloud Control SDK route (
src/app/api/discover/route.ts);src/aws/mcp.tsis the pure transform behind it. See CONTRIBUTING.md for the architecture.
The core idea: everything visual is derived from a data registry, not hardcoded. Adding a new AWS service is a single catalog entry — no UI changes.
Strata is open source. Contributions welcome.
Documentation
Strata ships its docs as part of the app, served at /docs, in two public
sections:
- User Guide (
/docs/guide) — how to use the app: building diagrams, importing IaC, validation, and saving/loading. - Architecture & Engineering (
/docs/architecture) — how it works internally: the service registry, domain model, persistence, MCP/IaC import, testing, and the roadmap.
The docs are authored as Nextra MDX under src/content/.
Quick Start
Strata is now a single app: npm run dev serves the product at / and the docs
at /docs.
npm install
npm run dev
- Product: http://localhost:3000/
- Docs: http://localhost:3000/docs
Saved diagrams persist in your browser (localStorage) — no external
infrastructure required, and it works on a read-only serverless host. Use
Export / Import JSON to move a diagram between browsers or share it. (The
server-side Repository + /api/graphs route remain for a future durable
backend, but the app no longer relies on them for saving.)
Other scripts
npm run build # production build (compiles the product and prerenders the docs)
npm run start # run the production build
npm run lint # lint
npm test # run the Vitest suite
Configuration
| Env var | Default | Purpose |
|---|---|---|
AWS_FLOW_REPOSITORY |
file |
Backend for the retained server tier (not used by browser-local save). |
AWS_FLOW_DATA_DIR |
.data/graphs |
Directory for the file-backed store (retained server tier only). |
AWS_FLOW_API_TOKEN |
(unset) | If set, the graph API requires Authorization: Bearer <token>. |
NEXT_PUBLIC_STRATA_HOSTED |
(unset) | Set to 1 on any shared/hosted deployment (see below). |
STRATA_INTEREST_WEBHOOK |
(unset) | Incoming-webhook URL to notify on "coming soon" interest (see below). |
"Coming soon" interest notifications
When NEXT_PUBLIC_STRATA_HOSTED=1, clicking a "coming soon" prompt's "I'd
use this" button POSTs a tiny { feature, note? } signal to the same-origin
/api/interest route (it's skipped locally — there's nothing to collect on a
single-user box). That route, with no datastore and no npm dependency:
- always logs the signal (so it shows in your host's function logs, e.g. Vercel), and
- if
STRATA_INTEREST_WEBHOOKis set, forwards a short message to that incoming-webhook URL. Use whatever gives you a notification with zero infra — a Slack or Discord incoming webhook (phone push), ntfy.sh (push), or a webhook→email relay. The body sends bothtext(Slack) andcontent(Discord), so one URL works for either. No credentials are involved, and the endpoint accepts only a small validated payload.
Live discovery & credentials
The Connect to AWS → Live scan flow runs server-side. With no credentials in the request, it uses the server process's default credential chain (env vars / shared profile / SSO / instance role) — fine for a single-user local run, where those are your own credentials.
On a shared/hosted deployment that ambient chain would be the operator's
account, so any visitor could enumerate it. Set NEXT_PUBLIC_STRATA_HOSTED=1
(at build and runtime) to disable the ambient fallback: each user must then
bring their own AWS credentials, entered in the modal and sent over HTTPS for a
single scan. Those keys are used in-memory only — never written to disk, logged,
returned, or saved into a diagram. Users should supply temporary, read-only
credentials (e.g. aws sts get-session-token, or an assumed ReadOnlyAccess
role). The credential-free Paste export tab remains available either way.
Project Structure
src/
aws/ Registry + domain model (the data foundation)
types.ts ServiceDefinition, ConfigField, RelationshipKind, scopes
registry.ts Aggregates catalogs; lookups, search, validateRegistry()
categories.ts Category + relationship presentation metadata
model.ts InfrastructureGraph / ResourceInstance / Relationship, validateGraph()
regions.ts AWS region reference list
rules.ts Architecture validation + best-practice rule suggestions
relationshipClasses.ts Edge visual encoding (colour/dash/arrowhead per class)
overlays.ts Analytical overlays (IAM-trust, network path, heat, reachability, tags) — pure
reachability.ts Policy-aware internet-reachability eval (subnets/routes/SG ports) — pure
review.ts "Explain & Clean" account review: cost + risk + safe-cleanup checklist — pure
cloudMap.ts Cross-cloud equivalence / migration (serviceId+provider rewrite) — pure
autofix.ts Deterministic fixes for validation findings (detect + apply) — pure
receipt.ts Change/audit receipt between two graphs (drift + cost + findings) — pure
tags.ts Tag collection + tint map + tag filter — pure
annotations.ts Notes / callouts / zones pinned to the canvas (excluded from rules/cost/IaC) — pure
dsl.ts Diagram-as-Code: graph ⇄ YAML round-trip — pure
iac.ts IaC import (CloudFormation + Terraform → InfrastructureGraph)
iacExport.ts IaC export (InfrastructureGraph → CloudFormation / Terraform scaffold)
mcp.ts Pure transform: DiscoveredResource[] → graph (mapDiscoveredToGraph)
discovery.ts Cloud Control descriptions → DiscoveredResource[] (no AWS SDK here)
services/*.ts Per-category service catalogs (networking.ts is the template)
canvas/ Pure geometry + containment layout (geometry.ts, layout.ts)
server/ Retained server tier (kept for a future durable backend)
repository.ts Repository interface
fileRepository.ts Default file-backed store
graphSchema.ts Zod schema for graph validation
auth.ts Optional bearer-token guard (AWS_FLOW_API_TOKEN)
index.ts getRepository() — backend selection via env
lib/
localStore.ts Browser localStorage save/load (the active persistence path)
api.ts Fetch client for the graph + discover APIs (discover is live)
app/
(product)/ The Strata canvas app, served at /
(docs)/ Nextra docs, served at /docs
api/graphs/ Graph REST Route Handlers (GET/POST, GET/PUT/DELETE by id)
api/discover/ Live discovery Route Handler (Cloud Control SDK, server-only)
components/ UI: Palette, Canvas, Inspector, CommandPalette
hooks/ Canvas state, rendering, interaction, undo/redo
content/ Nextra MDX docs (User Guide + Architecture), served at /docs
Adding a new AWS service
Append one ServiceDefinition to the matching catalog in src/aws/services/
(use networking.ts as the template) — the palette, colours, icons, inspector
form, and search pick it up automatically. See the
Service Registry docs.
Architecture
For the full design — registry schema, domain model, the registry-driven UI, the
swappable persistence layer, MCP ingestion readiness, and a candid list of gaps and
next steps — see the Architecture & Engineering docs
(src/content/architecture/).
Roadmap
What's shipped, the active follow-ups, and what's deferred until a backend exists: see ROADMAP.md.
Tech Stack
Next.js (App Router) · React · TypeScript · Tailwind CSS · Nextra (docs).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。