About Caterfli
Your software never needed replacing
We build the operator for the systems a business already owns, and we build it on the assumption that the AI driving it will sometimes be wrong.
Why we exist
The software works. Operating it is the job.
Most companies already bought the database, the API and the internal tools they need. Those systems work. What they lack is someone to drive them — so people spend their day clicking through admin screens to answer a question, raise an order or fix a record.
The usual answer is a migration: move everything into a new platform and retrain the team. We think that gets it backwards. Caterfli connects to what you already run — PostgreSQL, MySQL, a REST or OpenAPI service, an MCP server — learns what it can do, and hands it back as an agent your team can just ask.
Pointing AI at production data is only worth doing if it cannot break production. That is the part we spend most of our time on.
Your team asks
- Web chat
- Slack
- Telegram
Caterfli checks the call
create_refund- Tool exists
- Workspace owns it
- Person permitted
- Arguments valid
- Policy allows
- Approval gate
The software you already run
- PostgreSQL
- MySQL
- REST API
- OpenAPI
- MCP server
- Stays where it is
What we hold to
Principles you can check
None of these is an aspiration. Each is how the product behaves today, and the security model and documentation explain how.
The model is an untrusted planner
A language model proposes calls; it never executes one. Every proposal is checked independently for existence, ownership, permission, schema and policy before it reaches your systems.
Your data stays where it is
Caterfli holds a connection, not a copy. Nothing migrates and nothing gets rebuilt — disconnect it and your systems are exactly as they were.
A person decides what is destructive
Work that cannot be undone waits for a human, and an approval covers one call with the exact arguments that were reviewed. Never a capability.
No lock-in to a model provider
Anthropic, OpenAI, Google and OpenAI-compatible endpoints are configuration, not code. Switching does not change how your agents behave.
Limits written down in advance
Every allowance that can block you is on the pricing page before you sign up, and a lapsed subscription restricts features rather than deleting anything.
Say what we do not do
An approved deletion is still a deletion and a model can still be wrong. We would rather you knew that before connecting anything.
Working with us
A technical conversation, not a demo script
We will tell you if it does not fit
Tell us what you would connect first. We will say honestly whether it works today, whether it is on the list, or whether you would be better off waiting.
Try it before you talk to anyone
The free plan is the same product with smaller limits, and a demo connection lets you watch a real run before pointing anything at your own systems.
People read what you send
The contact form goes straight to our team, and a person reads every tester application — you hear back either way.
Security reports are welcome
Found something? Tell us before disclosing publicly. We will confirm receipt, keep you updated through the fix, and credit you if you would like that.
See what your software can do with an operator
Connect one system, read the agent drafted from it, and decide which of its tools stop for a person.