R&D · TRACK 03 / 05 · Shipped

An agent thatasks before it acts.

How the assistant calls tools by itself, what it is allowed to touch, and how you see and approve each step. All of it runs in one local process.

Read the notes View the code

What it does.
One system.

APPROVE · LIVE

A bounded turn loop

Each reply is a sequence of steps with a hard budget. The last step never runs tools: the model is told to answer from what it has.

6 STEPS

Permissions and approvals

Every tool has a permission level. Risky actions open an approval card in the app, and nothing runs until you allow it.

APPROVAL CARDS

Skills and sub-agents

Skills add focused instructions on demand. A sub-agent can take a task with its own budget and report back a summary.

SUB-AGENTS

A stop that works

Every step boundary checks the stop flag, so the Stop button and the kill switch end a turn cleanly.

KILL SWITCH

How it works

How it works.

Step 1

Read

The question is checked for anything that needs live data.

Step 2

Think

The model answers or asks for one tool call.

Step 3

Act

The tool runs under its permission, with approval if needed.

Step 4

Answer

Results come back and the model writes the reply.

Spec sheet.

01 / 05

Default steps

Default steps.

6 per question, configurable

Default steps
6 per question, configurable
Tool calls
Read from the visible reply, never from hidden thoughts
Approvals
Cards in the app for risky tools
Sub-agents
Nested loops with their own step budget
Events
Live server-sent events for every step

In short.
Agent loop.

Step-by-step tool calls with a hard step budget

Permissions per tool, with approval cards for risky actions

Skills and sub-agents for longer jobs

FROM IDEA TO RELEASE

Builders, not managers.

Say hello, fill in the blanks

Hi, I’m

and I want to talk about

.

Reach me at

Opens your email app, addressed to support@imutechz.in. Nothing is sent from this page.

or just email support@imutechz.in