Ailimu documentation

The problem

A FileMaker solution is usually the most complete record a business holds of itself, and among the hardest to interrogate. The answer to a question such as "which customers have invoices unpaid beyond sixty days, and what are those invoices attached to" is present in the data, but retrieving it means opening the correct layout, constructing a find, exporting the result and joining it to something else by hand. The question therefore goes to whoever knows the solution best, or it is not asked at all.

Providing an AI agent with a raw connection does not resolve this. An agent given a complete schema and a query language has to infer too much: which of several hundred table occurrences holds the live records, what the zz_ and zg_ calculation and global fields are for, and whether an empty column is empty because there is no data or because it was never adopted.

Ailimu resolves those questions during setup, under your review:

• It reads the real schema rather than a layout's view of it: true field types, declared keys, relationships, row counts, and the field comments already written in your solution.
• It determines what each table is for, whether core records, transactions, lookups, logs or join tables, and states its reasoning so you can correct it.
• It samples a bounded number of rows to identify columns that are never populated, and text fields holding only a handful of distinct values, so that search parameters carry meaning.
• You decide, field by field, what may be read, what may be searched, and what, if anything, may be changed.

The agent is left with a compact set of well described tools rather than a database it must reverse engineer. That is the entire premise: an agent given twenty good tools answers more reliably than one given two hundred bare ones.

What leaves your server

This is the first question any competent reviewer asks, so here is the direct answer.

Ailimu runs on infrastructure you control: your own network, a cloud server or a VPS. It is not a cloud service, it does not sync, and there is no copy of your database anywhere but your own server. But when the agent answers a question about your data, it has to see that data. The rows a tool returns are sent to the model, in that moment, to be reasoned about. This holds for every connector of this kind, including those offered by Claris and Ottomatic. Any product claiming otherwise is describing something that cannot answer questions.

What is sent to the model

Only two things, and only at the moment Claude calls one of your tools:

• The rows that tool returned. Only the fields you marked readable, and only for the records that matched the request.
• The names and descriptions of your tools, so Claude knows what it can ask for.

What stays on your server

Everything else. In particular:

• Your FileMaker account name and password. Read from the environment, never written to the config file, never shown in the browser.
• Any table or field you did not expose. It is not in any tool, so it cannot be asked for.
• Your database as a whole. Nothing is copied, exported, mirrored or indexed. Every tool call reads FileMaker live and returns only what that call asked for.
• Values sampled during setup. They are summarised into counts for you to see, then discarded. They are never saved.
• Your settings, tokens, wizard accounts and the activity log. All on the config volume of the server you run.

If your FileMaker Server is hosted somewhere else, Ailimu reads from it over HTTPS, the same way any FileMaker client does. What reaches the model does not change.

Two more requests can leave your server, and neither carries any of your data. Once a day Ailimu asks Docker Hub which versions of the Ailimu image exist, so the wizard can tell owners and editors when a newer one is out. When there is one, it also reads ailimu.net/releases.json, the headlines from the changelog, so the notice can say what the new version brings. Neither request says anything about your install beyond the address it comes from, as with any web request. Set UPDATE_CHECK=off to stop both.

Controls you have:

• Read-only is on by default. While it is on, no update tool and no script tool is generated at all, irrespective of what else is selected.
• Exposure is per field. Read means the value can come back; search means it can be filtered on. Everything else is invisible.
• Your FileMaker privilege set still applies. Ailimu connects as an account you create. Anything that account cannot see stays unseeable, whatever the config says. Issue it a restricted account.
• Everything that changes data is recorded. Every change and every script call goes into the activity log, because nothing else records what a script did. Reads can be recorded too, per connection. Search it all from Activity in the wizard.
• Container fields are never sent. They are excluded from every tool.

If your data may not reach a third-party model at all, then no connector of this kind is appropriate. That is a decision about the model, not about Ailimu.

How it fits together

Ailimu is one container, and it runs anywhere you can run Docker: a server on your own network, a cloud server, a VPS. It does not have to sit beside your database. It connects out to FileMaker Server over HTTPS, or to PostgreSQL over its port, wherever that database runs. Claude reaches Ailimu over HTTPS.

• Database on the same network. The simplest case: Ailimu reaches it directly, and nothing crosses the internet except Claude's tool calls and the rows they return.
• Database hosted elsewhere. Point Ailimu at the host's address, as any FileMaker client would. The data already travels between that host and your users; Ailimu is one more client of it, over HTTPS, signing in as the account you create. Ask your host to enable the Data API and OData for that account.
• Claris FileMaker Cloud is not supported yet. It signs API clients in with a Claris ID rather than a FileMaker account and password, which Ailimu does not do today.

Ailimu speaks MCP, the open standard for giving an agent tools. Claude connects to it the same way as to any other MCP server.

What it is not

• Not a chat interface. You bring the agent; Ailimu supplies the tools.
• Not a reporting tool. It answers questions live; it does not build dashboards.
• Not a sync engine. No copy of your data is made.
• Not a replacement for your privilege sets. It is another client, subject to them.

Where to go next

Start with Install. A first run takes about ten minutes: pull the image, set two environment variables, open the wizard, tick the tables you care about, publish, and paste one command into Claude.