Skip to main content
Glama
README.md
<img src="public/icon.svg" width="64" height="64" alt="Seedwall MCP icon">

# Seedwall MCP

Docs: **https://mcp.seedwall.suryansh.work**

[Seedwall](https://seedwall.suryansh.work) generative art inside Claude: hero and section backgrounds, OG images, banners, blog covers and wallpapers. Every image is a pure function of (look, preset, params, size, seed), so URLs are stable and safe to embed.

The repo has two parts, and both use the same MCP server:

- **A claude.ai connector**: a remote MCP server (Streamable HTTP) at `https://mcp.seedwall.suryansh.work/mcp`.
- **A Claude Code plugin** (`plugin/`): connects to that server and adds a `seedwall` skill. The skill reads your project's brand colours, picks a look, renders it, downloads the file into `public/` and wires it into your hero with responsive CSS, dark mode and a contrast check.

## Tools

| Tool | What |
| --- | --- |
| `list_looks` | Every look (generator) with its presets, plus themes and device sizes |
| `describe_look` | One look's params (colours with bg/fg/accent roles, density, text layer) and presets |
| `render` | Any look at any size. Returns SVG and PNG URLs, a studio link for editing, and a PNG preview the model can see |
| `seed` | A numbered seed (1–999999), or today's wallpaper |

All of them are read-only. The server is stateless and holds no data; rendering happens on Seedwall's public API (`/api/generators`, `/api/render`).

## Use it

**claude.ai**: Settings → Connectors → Add custom connector → `https://mcp.seedwall.suryansh.work/mcp`.

**Claude Code**:

```bash
claude plugin marketplace add Suryansh-Codes2209/seedwall-mcp
claude plugin install seedwall@seedwall
```

Then ask something like *"add a topographic hero background in our brand colours"*.

To use only the MCP server, without the skill: `claude mcp add --transport http seedwall https://mcp.seedwall.suryansh.work/mcp`.

## Develop

```bash
npm install
npm test                 # unit + in-process MCP protocol tests (2025 and 2026 protocol eras), Seedwall faked
npm run dev              # HTTP server on http://127.0.0.1:8787/mcp
npm run stdio            # stdio transport

# Live smoke test against a running Seedwall (e.g. `npm run dev` in ../Seedwall):
npm run build && SEEDWALL_URL=http://localhost:3000 npx tsx scripts/smoke.ts stdio
```

To point the plugin at a local server: `SEEDWALL_MCP_URL=http://127.0.0.1:8787/mcp claude --plugin-dir ./plugin`.

Environment:

- `SEEDWALL_URL`: the Seedwall instance to call (default `https://seedwall.suryansh.work`).
- `SEEDWALL_MCP_URL`: overrides the plugin's server URL.

## Deploy

The server deploys to Vercel as is: `api/mcp.ts` is the function, and `vercel.json` rewrites `/mcp` to it. Point the `mcp.seedwall.suryansh.work` domain at the project. It needs Seedwall's `/api/generators` and `/api/render` endpoints to be live.

## Privacy

The server stores nothing and has no accounts. Tool arguments (look, colours, size and so on) are forwarded to seedwall.suryansh.work to render the image, and standard hosting request logs apply. No personal data is collected.

## Licence

MIT

TDQS

A3.5/5.0

Scored across 4 tools

Disambiguation4/5

list_looks and describe_look are clearly split (all looks vs. one look's parameters), and render is a distinct parameterized generation tool. The only blurry boundary is seed, whose name reads like a render parameter and whose relationship to looks is unspecified, so an agent could hesitate between render and seed for a 'random image' request.

Naming Consistency3/5

list_looks and describe_look follow a clean verb_noun pattern, but render drops the object and seed is a bare noun that doesn't follow the verb convention at all. The names remain readable, but the set mixes conventions rather than holding one pattern.

Tool Count4/5

Four tools for a focused render/discovery service is lean but defensible: discovery (list/describe) plus two generation entry points. It is slightly thin, and seed arguably duplicates a mode that could live inside render, but nothing is bloated.

Completeness4/5

The core lifecycle is covered: discover looks and themes/sizes, inspect parameters and presets, then render to SVG/PNG with a studio edit link or grab a fixed/daily seed image. Minor gaps exist — no explicit standalone enumeration of sizes/themes, no way to combine a seed with a specific look, and no batch render — but agents can work around them.

Maintenance

ActivityMaintained
ResponsivenessOutdated data