Your own APIs · Tools
How do we turn an OpenAPI file into tools for an AI agent?
Give Katonic the OpenAPI or Swagger description of an API you already run. Each operation in it becomes a tool an agent can call, in the same gateway and under the same rules as your MCP servers. The API stays where it is, on your network, and Katonic calls it from inside your own installation.
How it is connected.
In Control Room, under Tools.
- 01Give the description. An address, pasted text or an uploaded file. A check reports errors, warnings and the number of tools before anything is created.
- 02Say how to sign in. The credential is stored encrypted and added to each call.
- 03Choose who may use it. Pick the teams, set a rate limit, and switch off any operation an agent should not have.
- 04Try it. Run a tool with real values, or as a dry run that shows the exact request and sends nothing.
What becomes a tool
Every operation in the description. A tool takes its name from the operation's id. Its inputs are the operation's parameters and the fields of its request body, with their types and which are required. You can rewrite a tool's description so an agent reads it the way you mean it.
What leaves your environment
Nothing, where the API is yours.Katonic calls the address you give. An API that someone else hosts receives what the agent sends it.
Policy
Each tool gets one of four answers: allow, allow and watch, hold for a person, or block. A policy can cover one tool or every tool from an API, for the whole organisation or one team. Where two policies meet, the stricter one wins. If the policies cannot be read, the call is blocked.
Approval
Once an agent is published, a tool that changes something waits for a person to approve it, until an administrator decides otherwise.
Caller identity
The signed-in person's id or email can travel with each call, in a header you name. The platform sets it. Nothing the model writes can replace it, so an agent cannot act as someone else.
Risk level
Each tool is rated by what it does: reading is low, creating or updating is medium, deleting is high.
Limits
Calls per minute for each API, 60 unless you change it. A time limit of 30 seconds for a call. Answers to read calls can be cached for up to a day, and are not unless you switch that on.
Health
Each API is checked every five minutes and shown as healthy, degraded, unreachable or sign-in expired, with the reason.
The API needs an OpenAPI 3 or Swagger 2 description. A SOAP service, a message queue or a file transfer needs a REST face in front first. Where an API returns a file, the agent is told its type and size and does not receive the file.
- 01MCP servers: the servers an agent can act through
- 02Katonic as an MCP server: offer an agent to other systems over MCP
- 03
Is an imported API turned into an MCP server?
No. Its operations are registered in the same tool gateway as your MCP servers, so an agent, a policy and the tool list treat both alike. An agent built on them can itself be offered to other systems over MCP.
Can an agent change data through an imported API?
Only where you allow it. Each operation is its own tool, so you can leave the reading ones on and switch the rest off, or hold them for approval.
Does the API have to be on the internet?
No. It can sit on your own network. Katonic calls it from inside your installation.