Give Your AI Assistant Real Answers: The ToolForte MCP Server

Ask a language model whether NL91ABNA0417164300 is a valid IBAN and it will say yes with great confidence. Change one digit and it will often still say yes. The model is pattern matching; the IBAN checksum is arithmetic. That gap is what the ToolForte MCP server closes: the assistant asks a tool that computes the answer, and repeats what the tool said.

MCP (the Model Context Protocol) is the plug standard the assistants agreed on. ToolForte runs a remote MCP server at https://toolforte.com/api/mcp (the older /api/mcp/mcp address keeps working). You add that address to your client once, and the tools appear in the assistant's tool list.

What is on the server

At the time of writing the server answers to 180 tools in six groups.

150 deterministic utilities. IBAN and VAT validation, VAT calculation, cron expression parsing, regex testing, JSON formatting and validation, UUIDs, hashes, timestamp conversion, unit conversion, Dutch test data (BSN, IBAN, BRP persons, UPA files) and the rest. These are the same functions the tool pages and the REST API use, so the three never disagree.

14 render tools. HTML or a public URL to a PDF, a screenshot of a page, a QR code as a PNG, PDF merge and split, images to PDF, PDF to Word and back, image resize, convert and compress, metadata stripping, and a page reader that returns a JavaScript-rendered page as Markdown. Each returns a hosted link that stays valid for 24 hours.

Four memory tools. memory_set, memory_get, memory_list and memory_delete give an assistant a place to keep a value between sessions, on your account. Free, unmetered, kept for 30 days on a free account and a year on Pro.

Three workflow tools. workflow_list shows the workflows saved on your account with the inputs each one wants, workflow_run runs one and returns every step's output, and workflow_propose turns a sentence like "resize these, then make a PDF" into a draft workflow. More on that in the workflows post.

Six AI tools. A scam message checker, an Excel formula writer, SQL from a question, code translation between languages, a prompt improver and an email writer. These are metered at 10 credits a call, with a small free daily allowance.

Connecting takes two minutes

Every client has its own page with the exact settings: Claude, Claude Code, Cursor, ChatGPT, VS Code, Windsurf and Gemini CLI. The shape is always the same: a remote server of type streamable HTTP at the address above, and optionally your API key in the X-API-Key header.

For a client that has no field for a remote URL, there is a bridge: npx -y toolforte-mcp runs a small local process that speaks stdio to the client and HTTP to us. It carries your key from an environment variable, which is also the answer for the claude.ai custom connector, which at the time of writing has no header field at all.

What it costs

Without a key, the 150 utilities and agent memory are free, up to 200 requests a day per client session, and each session also gets 3 renders a day to prove the render tools work. That is enough to try everything.

With a key, calls draw from the same balance as the REST API: one credit per utility call, 25 per render, 0 for memory. A free account has 1,000 credits a month; Pro has 7,500 for €19. When the balance is gone, the tool result says so in plain words, with a link, instead of returning a half answer. An assistant reads that and tells you; it does not silently degrade.

What changes in practice

I use this most in Claude Code while working on pension test data. "Generate ten BRP test persons born between 1955 and 1960 with a fixed seed" becomes a tool call, and the ten persons that come back are exactly the ten I get on the tool page with the same seed. The assistant did not invent a BSN; it received ten valid ones.

The second thing that changes is cron. I ask what 0 3 /2 1-5 does, the assistant calls the parser, and I get the next five run times rather than a paragraph that sounds right. The cron parser page does the same for people who prefer a form.

The third is the page reader. "Read this documentation page and summarise the auth section" fetches the rendered page as Markdown and hands it to the model. That is a render, so it costs 25 credits, and it only reaches public addresses.

Two honest limits

The assistant still has to decide to call the tool. Most of the time it does, because the tool descriptions say exactly what each one is for. Sometimes it answers from memory anyway, and then you are back to a guess. If a number matters, ask it to check with the tool by name.

And the server does what the tools do, nothing more. It will validate an IBAN; it will not look up the account holder, because nothing on ToolForte does that. The MCP page lists every tool with one line each, so you can see the edges before you connect.