VanAuto Campers

VanAuto Campers MCP

Read-only MCP server that lets an AI client search and view VanAuto camper listings, rentals and marketplace (akcesoria) offers. It talks to the public VanAuto API, so it needs no credentials and cannot change any data.

Public endpoint: https://mcp.vanauto.pl/mcp (Streamable HTTP, no auth) Health check: https://mcp.vanauto.pl/health (uptime-checked every 6h via GitHub Actions) MCP Registry: pl.vanauto/campers — discoverable by registry-connected clients.

Tools

Prompts

Clients that support MCP prompts get one-click flows:

Data is served in Polish (as on the site); prices are in PLN. Rental rows carry dailyRates (per-period rates) rather than a single price.

Connect a client

Remote (hosted)

OpenCode (~/.config/opencode/opencode.json):

{

"mcp": { "vanauto-campers": { "type": "remote", "url": "https://mcp.vanauto.pl/mcp", "enabled": true } } }

Claude / ChatGPT: add https://mcp.vanauto.pl/mcp as a custom connector.

Claude Desktop (older builds without native remote support):

{

"mcpServers": { "vanauto-campers": { "command": "npx", "args": ["mcp-remote", "https://mcp.vanauto.pl/mcp"] } } }

Local (stdio)

The package is npm-ready as `vanauto-mcp-campers`. Once it is published:

npx -y vanauto-mcp-campers

Until then, run it from a checkout with an absolute path to src/server.js (see mcp.config.example.json):

{

"mcpServers": { "vanauto-campers": { "command": "node", "args": ["/path/to/mcp-campers/src/server.js"], "env": { "VANAUTO_API_URL": "https://api.vanauto.pl", "VANAUTO_SITE_URL": "https://vanauto.pl" } } } }

Configuration

Development

npm install

npm start # stdio server npm run dev # local HTTP endpoint at http://127.0.0.1:8787/mcp (wrangler dev)

Checks

npm run check            # syntax check

npm run smoke # live tool checks against the configured API npm run smoke:protocol # MCP handshake + tools/list + tools/call over stdio npm run smoke:http # same over HTTP: MCP_URL=https://mcp.vanauto.pl/mcp npm run smoke:http npm run usage:report # anonymous tool-usage report from D1 (last 30 days)

Deployment

The server runs as a Cloudflare Worker (src/worker.js); the stdio entry (src/server.js) is for local clients. Both share the same tool factory in src/app.js.

npx wrangler login       # once per machine

npx wrangler deploy # deploys the Worker + mcp.vanauto.pl custom domain

Pushes to main deploy automatically via .github/workflows/deploy.yml, which needs two repository secrets:

The workflow runs npm run smoke:http against https://mcp.vanauto.pl/mcp after each deploy.

MCP Registry

The server is published to the official MCP Registry as pl.vanauto/campers (namespace verified via a DNS TXT record on vanauto.pl). To publish a new version after changing server.json:

# Signing key lives outside the repo, e.g. ~/.config/vanauto-mcp/registry-key.pem

PRIVATE_KEY="$(openssl pkey -in ~/.config/vanauto-mcp/registry-key.pem -noout -text | grep -A3 'priv:' | tail -n +2 | tr -d ' :\n')" mcp-publisher login dns --domain vanauto.pl --private-key "$PRIVATE_KEY" mcp-publisher publish

Keep version in server.json in sync with package.json; the registry rejects duplicate versions.

Usage analytics

The deployed Worker writes one anonymous row per tool call to Cloudflare D1 (tool_events): tool name, sanitized and truncated arguments, ok/error, duration, and a coarse client family parsed from the user agent (chatgpt, claude, opencode, ...). No IP addresses, tokens or user identifiers are stored.

npm run usage:report             # last 30 days

npm run usage:report -- --days 7

The report shows call counts per tool, client families, rental vs sale searches, and the most requested phrases, makes and camper types — useful market signal for what buyers ask AI about. Locally, wrangler dev writes events to a local database instead of the live one.

Behaviour notes

• Read-only. Every call is an HTTP GET against /public/...; there is no database access and no auth token.
• One public surface. Only mcp.vanauto.pl serves the MCP; the workers.dev route is disabled so the zone rate-limit rule cannot be bypassed.
• Caching. Catalog endpoints (makes, equipment) are cached in-process for an hour; searches and details for two minutes. The API itself also caches responses. On Workers, each isolate has its own short-lived instance of that cache.
• Pagination. search_campers returns up to 50 rows plus a nextCursor; pass it back as after for the next page. Marketplace search is page-based (page, perPage).
• Truncation. Long listing descriptions are trimmed (default 4000 chars, maxDescriptionChars to change), so one call cannot flood a model's context.
• Photos. Photo URLs are always included; includePhotos additionally attaches up to 6 (default 3) 400px WebP thumbnails (~25 KB each) as MCP image content for multimodal models.
• Tool failures come back as isError results with a readable message, e.g. a missing slug reports "No active sale or rental listing found with slug ...".