Custom tools for AI agents: when the ready-made ones are not enough

Alexander Krastev · 2026-10-19

Ready-made agents cover the common work. Every business also has one system nobody else uses — an old ERP, a supplier portal, a niche API. That is where custom tools for AI agents come in. AIK gives you four ways to add one, from no code to full control.

1. Any API, with its key — no code

Most services publish an API. Save its key once under Integrations → Other services (for example WEATHER_API_KEY) and an API agent can call it: you describe the request in words and the agent reads the answer. Good for a weather service, a CRM, a courier, a price list.

2. A program the agent writes for you

The coding agent writes a small Python program for a task — reshape a spreadsheet, compute a report — saves it, and runs it in a clean, closed environment with no network access. Running it is an action, so it waits for your OK first.

3. A script you already have

Have a script that already does the job? It can become an agent of its own: AIK runs it with the person's words as its input and reports what it printed. Adding script agents is limited to the platform's administrators, because a script runs on the server.

4. An MCP tool server

MCP is the open standard for giving AI agents tools. AIK's own tools are MCP servers, and on your own installation an agent can be connected to an outside MCP server over https — one per agent, every call treated as an action that waits for approval. The switch for this is off by default.

Whatever you add, the same rules apply

A custom tool does not get a back door: the agent still has its allow-list of recipients and sites, its budget, its approval step before acting, and its run log. On your own server the agent's tools can also be fenced into a sandbox that sees only that user's files.

More: secure AI agents · AI agent integrations.