by-ai-for-ai

Built by AI,
for AI.

Tools an AI built so AI agents can work on real systems safely, with a human in charge. Designed for the models you run yourself, including small local ones. Designed, written and tested live by Claude, working with vehoelite.

How we work

The human decides

What gets built, which devices may be touched, and where the lines are.

The AI shows its work

Dry run first. Risky changes need a confirm token. Everything is journaled, verified and undoable.

Tested on real equipment

Every write a tool claims to support has run against real hardware, not just a simulation.

Honest records

Changelogs keep the mistakes and how they were fixed. Trust comes from seeing both.

Made for local models

A small model on your own machine makes more mistakes than a frontier model. These tools are built so a mistake costs nothing: the tool does the careful part, and the model only has to follow the steps.

1Dry run by default

Every change is a preview unless the model asks for apply=true. It sees the exact plan, a field-by-field diff, warnings and the risk level. Nothing is sent.

2Two deliberate steps for risky changes

Critical changes need the confirm token from that dry run. It's tied to the exact plan and the device's current state, and it expires in 5 minutes.

3The tool checks its own work

After a change it reads the device back and reports applied, mismatch or no_effect. The model doesn't have to judge.

4Everything is undoable

Each change is journaled with the state before it. One call with the journal id puts it back.

It always says what's next

Replies carry a next step, and errors come with a hint and the tools that do work. A small model never has to guess.

One tool per area

Related actions share one tool instead of hundreds of tiny ones. The full router set is about 11k tokens, so give the model a 16k+ context.

Secrets stay out of the chat

Passwords are masked in every reply and read from files, never passed as arguments.

Projects

opus-tplink-mcp

MCP server for TP-Link Archer routers and RE extenders: dry run, verify, journal, undo. pip install opus-tplink-mcp

released
SSH tools

Safe remote work for agents, built the same way.

planned
Android tools

Drive a phone and confirm every step.

planned

Why

A lot of people expect AI and people to end up on opposite sides. These tools are a small argument for the other outcome. It's better to build things together than to spend that energy pulling apart.