About software-design
software-design is an MCP server for coding agents and developers, offering references on design principles and GoF patterns, pattern scaffolding, and heuristic code-smell checks. It leaves architectural judgment to the model rather than treating heuristics as verdicts.
Key features
- 01Browse principles and pattern catalogs
- 02Explain principles and pattern concepts
- 03Generate pattern pseudocode skeletons
- 04Detect code smells heuristically
- 05Provide design review and pattern prompts
Use cases
- 01Look up SOLID principles or GoF patterns
- 02Scaffold a pattern for code
- 03Spot long methods or deep nesting
- 04Review code using design principles
MCP support
It exposes tools, resources, and prompts to MCP clients: tools look up concepts, scaffold patterns, and check code smells; resources provide reference material, and prompts support design reviews and refactoring.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
4Saved tool list · updated 3w ago
| Parameter | Type | Required | Description |
|---|---|---|---|
| kind | enum | No | Filter by kind. "principle" = all principles; "solid"/"oop" narrow to the SOLID five / the four OOP pillars; "pattern" = all GoF patterns. Default "all".all | principle | solid | oop | pattern | creational | structural | behavioral |
Details
From the project READMEmcp-software-design
An MCP server that teaches and helps apply software-design guidance — the SOLID principles, the OOP pillars, DRY / KISS / YAGNI / meaningful naming / clean code, and the 23 Gang-of-Four design patterns — plus pattern scaffolding and heuristic code-smell detection.
It's the companion to
mcp-udacity-commit: same stack (TypeScript, the MCP
SDK, stdio transport), same shape (pure logic modules + thin server wiring).
Install
Register it with Claude Code — one line, nothing to clone:
claude mcp add software-design -- npx -y mcp-software-design
Or in an MCP client config:
{
"mcpServers": {
"software-design": {
"command": "npx",
"args": ["-y", "mcp-software-design"]
}
}
}
Why this exists — and its one honest caveat
The commit server can lint: "subject ≤ 50 chars" is objectively checkable. Design principles and patterns are not like that — "does this violate SRP?" or "should this be a Factory?" are judgment calls, not lint rules.
So this server does not pretend to grade your architecture pass/fail. Instead it does the parts that are genuinely reliable, and hands the judgment to the model:
| Capability | Primitive | What it gives you |
|---|---|---|
| Reference | resources + explain_concept | Authoritative, consistent definitions so the model cites the same thing every time. |
| Scaffolding | scaffold_pattern | A language-agnostic skeleton of a pattern's participants. |
| Smell heuristics | check_smells | A few genuinely-checkable proxies (long method, deep nesting, …) — hints, never verdicts. |
| Explain / apply | review_design, apply_pattern prompts | Prime the model to review or refactor, grounded in the tools + resources above. |
Design analysis is a judgment call, so the "explain/apply helper" is exposed as MCP prompts (which drive the client's model) rather than server code pretending to understand your snippet.
Tools
list_catalog{ kind? }— list concepts, optionally filtered (principle|solid|oop|pattern|creational|structural|behavioral).solid/oopnarrow to the SOLID five / the four OOP pillars.explain_concept{ name }— full guidance for one principle or pattern (intent, when-to-use, trade-offs, participants). Accepts a slug, name, or alias ("SRP","open-closed","pubsub").scaffold_pattern{ pattern, names? }— pseudo-code skeleton for a GoF pattern;namesoptionally renames roles to your domain ({ "Product": "Notification" }).check_smells{ code, …thresholds? }— heuristic scan for long method, large class, long parameter list, deep nesting, duplication, and large file. Each finding names the principle it hints at plus a suggested refactor. All thresholds are overridable per call.
Resources
design://principles— SOLID, OOP pillars, DRY, KISS, YAGNI, meaningful naming, and more.design://patterns— the 23 GoF patterns, grouped creational / structural / behavioral.design://smells— whatcheck_smellsdetects, its thresholds, and its caveats.
Prompts
review_design{ code, focus? }— review a snippet against the principles/patterns, grounded incheck_smells+ the resources.apply_pattern{ pattern, code }— refactor a snippet to apply a named pattern (and first judge whether it even fits).
Build from source
For local development, or to run a local checkout instead of the published package:
npm install
npm run build # compiles src → build
npm start # runs the stdio server
npm test # builds, then runs the unit tests
npm run test:client # end-to-end check against the built server
Then register it the same way as Install above — both the
claude mcp add command and the MCP-client-config form work — but point at
your local build instead of npx:
claude mcp add software-design -- node /absolute/path/to/mcp-software-design/build/index.js
Layout
src/
catalog.ts # principles + 23 GoF patterns (data + lookup + markdown)
smells.ts # pure, testable code-smell heuristics
scaffold.ts # renders a pattern's participants into a skeleton
index.ts # MCP wiring: resources, tools, prompts
test/
catalog.test.mjs # catalog lookup + scaffolder
smells.test.mjs # smell detectors (incl. string/comment edge cases)
The src/*.ts logic modules are pure and side-effect-free, so they're unit
tested directly against the compiled output — the server (index.ts) is only
thin wiring on top.
License
MIT
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.