beds24-mcp-server

MCP server + CLI + SDK for the Beds24 API. Two complementary layers over the same source of truth:

1. Semantic search — vector index over the cited markdown docs (beds24-search). Answers "how does pricing propagate?", "what are channel source IDs?".

2. Schema validation — resolves apiV2.yaml and validates draft payloads (beds24-validate). Answers "what's wrong with my POST /bookings payload?".

The markdown facts in knowledge/ + knowledge/apiV2.yaml are the source of truth. The .beds24/ vector index is a regenerable cache (bun run index). The SDK (src/sdk/) has zero MCP dependency so it can be imported from other repos (data-plattform, workflows, etc.).

Install

Global (recommended — harness-available everywhere)

npm install -g beds24-mcp-server

This publishes the beds24-mcp-server command. Then auto-configure your harness(es):

beds24-mcp-server setup        # detects Claude Code / Cursor / Windsurf / VS Code, writes their MCP config

That's it — setup writes the config, runs bun install (if needed), and builds the vector index. Restart your harness; the tools appear.

Run beds24-mcp-server setup --dry-run to preview the writes without touching anything, or --harness claude --harness cursor to pick specific ones. Use --skip-index if you want to build the index later.

Local / from source

bun install
bun run index      # build the vector index from knowledge/*.md (one-time, ~30s)
# or let it auto-index on first server start

To configure harnesses from a source checkout:

bun run setup      # same detection + config writing as the global command

Related MCP server: one-mcp

Using the SDK from another repo

The SDK has no MCP dependency — just point it at the facts + yaml:

import { Beds24Validator, Beds24Search } from "beds24-mcp-server/sdk";

const validator = Beds24Validator.create({ factsDir: "path/to/knowledge" });
const result = await validator.validate("POST /bookings", "request", payload);
// result.valid, result.errors — feed errors back to your LLM to fix the call

This lets you run schema validation from a dagster asset, an inngest function, or a CLI — no MCP server needed.

Using the client

A dependency-free HTTP client (global fetch, Node 18+) with 24h-token auth, automatic token refresh on 401, request validation, and credit-limit tracking:

import { Beds24Client } from "beds24-mcp-server/client";

const client = new Beds24Client({ apiKey: "...", propKey: "..." });
const { data, credits } = await client.request("GET /bookings", {
	arrival: "2026-08-01",
	departure: "2026-08-05",
});
// credits.remaining / credits.resetsIn — the 5-minute rate-limit window

Every METHOD /path in apiV2.yaml is reachable via client.request(key, body). Request bodies are validated against the schema before sending (fail fast, save a credit). Throws Beds24Error with status, code, retryable, and creditsRemaining.

Configure (your harness)

All harnesses speak the same MCP JSON shape; only the config file location differs.

The shared block (paste into command / args below). Prefer the global command if you installed via npm — it survives reinstalls and doesn't hard-code a checkout path:

{
  "command": "beds24-mcp-server",
  "args": ["serve"]
}

Falling back to a source checkout (replaces /ABSOLUTE/PATH/to/beds24-mcp):

{
  "command": "bun",
  "args": ["run", "/ABSOLUTE/PATH/to/beds24-mcp/src/server.ts"]
}

Run beds24-mcp-server setup (or bun run setup) to write these files automatically — no hand-editing needed.

Claude Code

Project-level .mcp.json (committed, shared with the team) or user-level ~/.claude/.mcp.json (just you):

{
  "mcpServers": {
    "beds24": {
      "command": "bun",
      "args": ["run", "/ABSOLUTE/PATH/to/beds24-mcp/src/server.ts"]
    }
  }
}

Restart Claude Code. Tools appear in every session.

Cursor

Global ~/.cursor/mcp.json (all projects) or project .cursor/mcp.json:

{
  "mcpServers": {
    "beds24": {
      "command": "bun",
      "args": ["run", "/ABSOLUTE/PATH/to/beds24-mcp/src/server.ts"]
    }
  }
}

Open Cursor Settings → MCP; the server should show green. If not, click Restart all servers (it must resolve bun on your $PATH — launch Cursor from a shell or add the bun path to the config's env).

Windsurf

~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "beds24": {
      "command": "bun",
      "args": ["run", "/ABSOLUTE/PATH/to/beds24-mcp/src/server.ts"]
    }
  }
}

Reload the window / restart Windsurf to pick it up.

VS Code (Copilot)

Project-level .vscode/mcp.json:

{
  "servers": {
    "beds24": {
      "command": "bun",
      "args": ["run", "/ABSOLUTE/PATH/to/beds24-mcp/src/server.ts"]
    }
  }
}

Note: VS Code uses servers (not mcpServers). Restart VS Code after editing.

Any other harness (OpenCode, goose, …)

The shape is the same — look in your harness's settings for "MCP servers" and register a server named beds24 with the command / args block above.

Troubleshooting

• "bun not found" — the harness doesn't inherit your shell $PATH. Launch it from a terminal, or add the bun dir to the server's env (e.g. "env": { "PATH": "/Users/you/.bun/bin:/usr/bin:/bin" }).

• No tools after restart — the server auto-indexes on first run and logs to stderr. Open the harness's MCP/output panel and look for [beds24] MCP server connected on stdio.. If it shows an error, re-run bun install && bun run index in the repo.

• Stale facts — after updating knowledge/, run bun run index (or just delete .beds24/ and restart; it rebuilds).

Tools

Source layout

.
├── knowledge/                                    # facts + OpenAPI spec (source of truth)
│   ├── index.md, api-v2/, pricing/, system-logic/ ...
│   └── apiV2.yaml
├── src/
│   ├── sdk/               # REUSABLE TS SDK (no MCP deps — import from other repos)
│   │   ├── index.ts       # re-exports
│   │   ├── db.ts          # libsql store + sqlite-vec cosine index
│   │   ├── embed.ts       # local embedding model (Xenova/all-MiniLM-L6-v2)
│   │   ├── chunk.ts       # markdown splitter (heading-aware, keeps citations)
│   │   ├── indexer.ts     # walk facts → section chunks → embed → store
│   │   ├── search.ts      # vector search + section lookup
│   │   ├── schema.ts      # parse apiV2.yaml → resolve $ref/allOf/oneOf
│   │   └── validate.ts    # draft payload → structured LLM-friendly errors
│   ├── server.ts          # thin MCP wrapper over the SDK (MCP deps only here)
│   └── cli.ts             # thin CLI wrapper over the SDK (`beds24-mcp-server index|status`)
├── .beds24/               # generated vector index (gitignored)
└── package.json

Indexing strategy (important)

The facts are already split by statement + source + date. Do NOT shred into fixed-size chunks — that destroys citations and mixes unrelated facts ("vector soup").

Instead, split at section (##) / subsection (###) boundaries. Each index entry = one section, storing its heading path, full text (citations inline), source file, and line range. A query lands precisely on the relevant cited section.

Structured-table content (the schemas-*.md files and version-reference.md) is better served by the schema/validate tools (exact lookup) than by vector search — route precise field questions there, fuzzy "how does it work" questions to search.

Refresh

When facts change: bun run index rebuilds the index. The server auto-indexes on startup if .beds24/ is missing.