How to build a FileMaker MCP server - connect AI agents to your Claris data

"Can I just ask Claude questions about our database?" is now one of the most common things clients ask us. The honest answer, as of 2026, is yes - and you don't need to write any code to do it. The mechanism is MCP, the Model Context Protocol: an open standard that lets an AI assistant securely reach into an external system, look things up and take actions. A FileMaker MCP server is simply an MCP endpoint that exposes your Claris FileMaker tables and scripts as tools an AI agent can use.

This guide is the version we'd give a client: what an MCP actually is, the two ways to build one for FileMaker, the prerequisites nobody mentions until you hit them, a step-by-step for the fast route, and - importantly - the FileMaker 2026 features that decide whether your MCP gives sharp answers or confident nonsense. As a working UK Claris Partner, our bias is towards what holds up in production, not what demos well.

What is an MCP server, and why point one at FileMaker?

An LLM on its own only knows what it was trained on and what you paste into the chat. MCP is the plumbing that lets it go and fetch live information, or perform an action, from a system it's been given access to. The AI client (Claude Desktop, Cursor, an in-house agent) speaks MCP; the MCP server translates those requests into operations against your data. For FileMaker, that means an assistant can:

• Search and read records across the tables you expose - "which open jobs are overdue for the North region?"
• Create and update records - logging a call, raising a task, updating a status.
• Run FileMaker scripts you've chosen to make available - triggering an existing business process rather than reinventing it.

Done well, this turns your FileMaker system into something staff can query in plain English, and something your other AI tools can build on. Done carelessly, it's an open door into your database - which is why the security section below matters as much as the setup. For the broader picture of what AI genuinely delivers inside Claris today, read AI & LLMs inside FileMaker - what actually works in 2026.

Two ways to build a FileMaker MCP: Claris MCP vs a custom server

There are two credible routes, and most organisations should start with the first:

• Claris MCP (no-code, official) - Claris ships a hosted MCP service. You connect it to a hosted FileMaker file through a web console, pick your tables and scripts, and it generates a configuration snippet for your AI client. Zero custom code, and it's the right choice for the vast majority of use cases.
• A custom MCP server (advanced) - you write your own MCP server (typically in Node.js or Python) that talks to FileMaker over the FileMaker Data API or OData, and expose exactly the tools you want. More work, but total control over auth, shaping and which operations are allowed.

Claris MCP first shipped on the FileMaker Server 22.0.2 line (late 2025). It runs perfectly well there - but FileMaker 2026 adds the schema-context and reliability features that make it worth building on, which we cover further down.

Before you start: the prerequisites

Almost every failed first attempt we see comes down to a missing prerequisite. For Claris MCP you need all of the following in place:

• FileMaker Server 22.0.2.204 or later, hosting the file you want to expose.
• Data API and OData both enabled in the FileMaker Server Admin Console - Claris MCP uses both under the hood.
• Extended privileges - the FileMaker account used for the connection needs both the fmrest and fmodata extended privileges, plus access to the specific tables and fields you intend to expose.
• A Claris ID and team-manager rights in your Claris Studio team, so you can access the Claris AI Workspace.
• A known 22.0.2 gotcha - on that server version, value list access should be disabled for the connection to avoid MCP errors. Worth knowing before you spend an hour debugging.
• Node.js 22+ on any machine running a command-based MCP client such as Claude Desktop.

If your FileMaker isn't hosted somewhere with the Data API and OData exposed securely, that's the first thing to fix - our managed FileMaker hosting handles that setup, and a Health Check will confirm your server is on a version and footing that supports this.

Step by step: building your MCP with Claris MCP

The no-code route runs through the Claris AI Workspace console. In outline:

1. Sign in at aiworkspace.platform.claris.com with your Claris ID and create a context - a container for one or more connections.
2. Add a connection. Enter your FileMaker Server domain, the exact database (file) name, and an account's credentials. Claris then shows the files that account can reach.
3. Select tables. Choose the tables you want the AI to be able to query and act on. Only what you tick is exposed - be deliberate.
4. Select scripts. Do the same for scripts. This is how you let an agent trigger existing business logic (e.g. "create invoice") rather than manipulating raw data.
5. Create the connection. Claris tests it and, if it succeeds, generates a set of MCP tools mapped to your tables and scripts.
6. Generate the configuration snippet. This is a small JSON block containing a URL and a token. You can use a direct SSE configuration, or a command-based "proxy" configuration for CLI clients. Treat the token as a secret - it grants access to everything in the context.
7. Connect your AI client. In Claude Desktop, for example, open Settings → Developer → Edit Config and paste the proxy configuration (it uses npx mcp-remote, hence the Node.js requirement), then restart the app.
8. Test it. Ask the assistant "Am I connected to Claris MCP?" and then "What FileMaker database am I connected to?". Correct answers mean you're live - now try a real question against your data.

That's the whole loop. The hard part isn't the clicks - it's deciding what to expose and making sure the data behind it is clean and well-described. Which is exactly where FileMaker 2026 earns its place.

How FileMaker 2026 makes your MCP dramatically better

You can run Claris MCP on the 2025 server line, but FileMaker 2026 (released June 2026) adds the features that separate an MCP that gives sharp answers from one that confidently makes things up:

• Field annotations for AI - the Advanced Options for Field dialog now has a proper annotation that describes a field for an AI model. It goes into the Data Definition Language (DDL) the model reads - and stays out of your human-facing field comment. The new FieldAnnotation() function reads it back. Behaviour to internalise: annotate no fields in a table and the DDL exposes all of them; annotate even one and the DDL exposes only the annotated fields. That gives you fine-grained control over exactly what schema an MCP client sees.
• Richer OData metadata - because Claris MCP leans on OData, this matters a lot. In 2026 the OData metadata now includes field comments, your AI annotations, field options and scripts (which can be executed by script ID). Any MCP tooling reading that metadata gets far more context about what your data means and what it can do.
• Self-hosted AI Model Server with RAG - if you'd rather not send data to a commercial provider, the AI Model Server now has its own lifecycle, runs embedding and text-generation workloads concurrently, and supports retrieval-augmented generation over your own text and PDF documents. That pairs naturally with an MCP that keeps data in-house.
• Google Gemini support - added as a provider for text generation and embeddings, alongside the existing options, so you can choose the model that suits your use case and budget.
• Self-healing services - the Data API, OData and Web Publishing Engine now restart automatically after a crash. Since those services power your MCP, this is the difference between an agent that's reliably available and one that silently goes dark at 3am.
• More OData throughput - FileMaker Server 2026 can run multiple script-engine processes, so heavy OData consumers and scripted work no longer starve each other. An always-on AI agent hitting your data is exactly the kind of consumer that benefits.
• Admin API automation - new endpoints manage the AI Model Server (start/stop, models, settings) and Standby Server, which helps if you're standing up this infrastructure as code.

In short: 2026 is the release where FileMaker stopped treating AI as a bolt-on and started feeding models real context about your schema. For the full rundown of the release, see FileMaker 2026 - what's new, and should you upgrade?

The custom route: building your own MCP server

Claris MCP covers most needs, but a custom MCP server makes sense when you need tool behaviour Claris MCP doesn't offer - bespoke auth, data shaping or validation, aggregating multiple systems behind one agent, or exposing carefully-scoped composite operations rather than raw table access. The pattern is straightforward:

• Write an MCP server (Node.js or Python are the common choices) using an MCP SDK.
• Have it call FileMaker over the Data API or OData - the same secure HTTPS endpoints Claris MCP uses.
• Define each tool deliberately: what it does, what it accepts, and what it's allowed to touch. This is where you enforce your own guardrails.
• Register the server with your AI client just like any other MCP server.

This is more work and more responsibility, but it's the right call for teams with strict governance or a genuinely custom agent. If you're weighing this against off-the-shelf automation platforms, our comparison of Claris Connect vs n8n vs Zapier covers where each fits.

Security, governance and UK GDPR

An MCP gives an AI agent a real door into your database, so treat it with the same care as any integration - more, arguably, because the caller is non-deterministic. The essentials:

• Least privilege - create a dedicated FileMaker account with a tightly scoped privilege set, not an admin login. Expose only the tables, fields and scripts the use case genuinely needs.
• Protect the token - the generated configuration token grants access to everything in the context. Store it as a secret, rotate it if exposed, and don't commit it to a repo.
• Mind data residency and encryption - the Data API and OData carry your data over HTTPS; your hosting, encryption-at-rest and residency posture still applies.
• Log and review - keep audit trails of what the agent reads and writes, and document it in your UK GDPR record of processing if personal data is in scope.

We run exactly these checks on every engagement - the full list is in our FileMaker Security & UK GDPR checklist, and it's part of every Health Check.

Our take: who should build what

• Most teams - use Claris MCP, on FileMaker 2026, with well-annotated fields and a tightly scoped account. It's fast, supported, and covers the "let staff ask our system questions" use case cleanly.
• Regulated or high-governance shops - consider a custom MCP server so you own the auth and can expose composite, validated operations rather than raw tables. Pair it with the self-hosted AI Model Server if data can't leave your estate.
• Still on FileMaker 19 or the 2025 line - you can start with Claris MCP today, but plan the move to 2026 for the annotation and OData-metadata gains. Get independent advice on the licensing first.

The short answer

Building a FileMaker MCP server in 2026 is genuinely a no-code job for most teams: enable the Data API and OData, wire up Claris MCP, expose only what you need, and connect Claude or your agent of choice. The quality of the result, though, comes from the unglamorous work around it - clean, well-annotated data on FileMaker 2026, a least-privilege account, and a clear view of what you're letting an AI touch. Get those right and you've turned your FileMaker system into something your whole organisation - and its AI tools - can actually talk to.

See also - AI & LLMs inside FileMaker, FileMaker 2026 - what's new, and FileMaker Security & UK GDPR.