Wafeq MCP Server

ohneben's Wafeq MCP

---

License & Checks

MCP Registries

Run your Wafeq books in plain language from AI assistants like Claude, Cursor, and any other MCP client.

This Model Context Protocol server exposes the Wafeq Public API — all 251 endpoints, generated straight from the OpenAPI spec into MCP tools, plus two hand-written ones. Every tool carries a safety category (🟢 read-only / 🟡 write / 🟠 state change / 🔴 irreversible or destructive) so your assistant knows what an action does before it calls it — including the difference between saving an invoice and filing it with a tax authority, which no CRUD-shaped wrapper can tell you. It runs over stdio (Claude Desktop and other local launchers) or Streamable HTTP (hosted in Docker), and ships with retries, client-side rate limiting, request timeouts, idempotency keys, multipart upload and binary PDF handling so it holds up against a live book.

Why you'll want this

Some MCP servers just forward an API. This one is built to be safe to hand to an LLM and easy to run against real accounting data:

How it compares

*Generic OpenAPI→MCP wrappers turn any spec into MCP tools. They can reach the same endpoints, but treat every operation identically — and against Wafeq's spec specifically they inherit the read-only-required-field problem described in MIGRATION.md. "➖" = varies by tool / not guaranteed.

What you can do

Once it's connected, ask your assistant things like:

• "What was our profit and loss for the first half of this year?"
• "Show me every unpaid invoice older than 30 days, with the customer name."
• "Create a draft invoice for Acme Ltd for 3 days of consulting at €800/day."
• "Download invoice INV-2026-014 as a PDF."
• "Which account did the €7,000 transfer in January post to?"
• "Attach this receipt to expense EXP-118."
• "Reconcile the bank statement lines for March against the ledger."
• "Convert quote QUO-31 to an invoice and record the payment."

How it works

Claude / Cursor / any MCP client  ──MCP──►  this server  ──HTTPS──►  Wafeq API (your organization)

At startup the server parses the bundled OpenAPI spec into MCP tools — resolving $refs, guarding against recursive schemas, and stripping server-assigned (readOnly) fields out of request bodies — tags each tool with its safety category, verifies which Wafeq organization the credentials belong to, and then injects your credential on every outgoing request. Your key stays in the server's environment; the model never sees or handles it.

Requirements

• A Wafeq organization with API access — either a private API key (Wafeq → Settings → Developer → API keys) or an OAuth2 access token. See Get your API credentials.
• Docker (Docker Desktop on macOS/Windows) for the quick start below — or Node.js ≥ 20 to run from source.

Quick start (Docker)

1. Add your credentials. Copy the example config and fill it in:

cp .env.example .env

Then edit .env and set WAFEQ_API_KEY. If the server will be reachable beyond localhost, set MCP_SHARED_TOKEN to a long random string as well.

2. Start the server:

docker compose up -d --build

docker-compose.yml binds to 127.0.0.1:8765 only, so the server is reachable from your machine but not from the network.

3. Confirm it's running — and that it's pointed at the right books:

curl -s http://localhost:8765/health

{
  "status": "ok",
  "server": "wafeq-mcp",
  "version": "2.0.0",
  "tools": 253,
  "organization": {
    "status": "ok",
    "id": "org_...",
    "name": "Your Company FZCO",
    "base_currency": "EUR",
    "country": "AE"
  },
  "auth_required": false
}

Check the name field. That is the organization your key writes to. If it isn't the company you expected, stop and fix the key before doing anything else. /health answers 503 and "status": "degraded" when the credentials can't be verified.

4. Point your MCP client at it: http://localhost:8765/mcp (Streamable HTTP).

Remote endpoints are added to Claude as a custom connector (Settings → Connectors), or bridged locally with mcp-remote. For the bridge, add this under mcpServers in your client config and restart the app completely:

{
  "mcpServers": {
    "wafeq": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:8765/mcp",
        "--header", "Authorization: Bearer YOUR_MCP_SHARED_TOKEN"
      ]
    }
  }
}

(Drop the --header line if you left MCP_SHARED_TOKEN empty.)

Prefer a ready-made image?

Every release publishes a ready-to-run image to the GitHub Container Registry, so you can skip the local build entirely:

docker run -d --name wafeq-mcp -p 127.0.0.1:8765:8765 --env-file .env \
  ghcr.io/ohneben/wafeq-mcp:latest

Pin a version (:2.0.0) rather than latest if you want releases to be something you opt into.

Install from the MCP Registry

The server is published to the MCP Registry as io.github.ohneben/wafeq-mcp, so registry-aware clients can install it by name. The registry entry launches the image over stdio — see Run the container over stdio for the equivalent hand-written config.

curl -s "https://registry.modelcontextprotocol.io/v0.1/servers/io.github.ohneben%2Fwafeq-mcp/versions/latest"

Get your API credentials

Private API key (most people): in Wafeq, go to Settings → Developer → API keys and create a key. It is scoped to one organization. Put it in .env as WAFEQ_API_KEY; the server sends it as Authorization: Api-Key <key>.

OAuth2 app: if you have an access token from a Wafeq OAuth2 app, put it in .env as WAFEQ_ACCESS_TOKEN instead. The server switches to Authorization: Bearer <token> automatically. Set WAFEQ_AUTH_SCHEME only if you need to force one scheme while both variables are present.

Configuration

All configuration is environment variables. Everything except the credential has a working default.

Too many tools?

251 tools is a lot. The full catalogue is about 0.5 MB of JSON (~133k tokens) on tools/list, and some hosts get slower or less accurate with that many. Two things help.

The schemas are already compacted. Wafeq's spec renders every enum's values into its description as well as into enum — the currency list alone is ~4 KB, inlined at 203 places. The generator collapses those single-member allOf wrappers and drops the duplicated bullet lists, which takes ~29% off the payload without removing a single allowed value.

Narrow the catalogue if you still want it smaller — no code changes needed:

WAFEQ_TOOL_GROUPS=invoices,bills,contacts,payments,reports,accounts,items,tax-rates

The two hand-written tools are always available, so nothing becomes unreachable — anything you filter out can still be called through wafeq_request.

Tool safety categories

Every tool's description opens with a banner, and every tool carries the matching MCP annotations. Counts are for the bundled spec (251 generated + 2 hand-written = 253).

The three 🔴 groups all set destructiveHint: true, so a host that honors annotations stops and asks before any of them — not just before deletions. Filing an invoice with a tax authority is at least as consequential as deleting one, and unlike a deletion it reaches outside your organization.

Print the live catalogue any time, without credentials:

npm run list-tools

Coverage

"Write" includes the two 🟠 state-change tools. Areas map to Wafeq resources as follows — Sales: invoices, simplified invoices, quotes, credit notes, payments, payment requests · Purchasing: bills, purchase orders, debit notes, expenses, beneficiaries · Banking: bank accounts with their ledger and statement transactions · Ledger & reporting: accounts, manual journals, journal line items, the four reports, tax rates, amortizations, revenue recognitions · Payroll: payslips, employees · Master data: contacts, items, units of measure, warehouses, projects, cost centers, branches, custom fields.

Two hand-written tools

Everything above is generated. Two tools are written by hand:

• wafeq_account_ledger (🟢) — journal line items with their real transaction date. Wafeq's /journal-line-items/ rows carry created_ts (when the row reached Wafeq), which is regularly a different month from the transaction, and no date field at all. This tool recovers the date using the endpoint's own date_after/date_before filters, which do operate on the transaction date. It probes a month at a time and only splits into day queries where rows exist, so quiet periods cost one request each; the result reports requests_made.
• wafeq_request (🔴) — the escape hatch: any method, any path, plus query, body and headers. It is the fallback for anything the bundled spec misses, not the primary interface. Categorized destructive because its effect can't be known in advance.

Run from source (stdio, no Docker)

npm ci
npm run build

Then register it with your MCP client. For Claude Desktop, add to claude_desktop_config.json:

{
  "mcpServers": {
    "wafeq": {
      "command": "node",
      "args": ["/absolute/path/to/Wafeq MCP/dist/index.js"],
      "env": {
        "MCP_TRANSPORT": "stdio",
        "WAFEQ_API_KEY": "your-key-here"
      }
    }
  }
}

For Claude Code:

claude mcp add wafeq --env WAFEQ_API_KEY=your-key-here -- node /absolute/path/to/dist/index.js

Run the container over stdio

You can also let your client launch the published image directly, with no HTTP server and no local build:

{
  "mcpServers": {
    "wafeq": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT=stdio",
        "-e", "WAFEQ_API_KEY",
        "ghcr.io/ohneben/wafeq-mcp:latest"
      ],
      "env": {
        "WAFEQ_API_KEY": "your-key-here"
      }
    }
  }
}

MCP_TRANSPORT=stdio is required here: the image defaults to the HTTP transport.

Useful scripts:

Keeping the spec current

Tools are generated from spec/wafeq-public-api.json at startup — there is no code generation step and no hand-written tool list. Drop in a newer OpenAPI document (JSON or YAML), rebuild, and new endpoints become new tools. See spec/README.md for where the bundled copy came from and what to re-check after an update.

Documentation truncated — see the full README on GitHub.