# What AgentCode Does

AgentCode is a local AI work assistant for one computer.

The first version is a private beta for a trusted Mac user. It is meant to help with files, folders, reports, scripts, and code tasks on the Mac where it is opened.

## What It Does In V1

- Opens through one visible model in OpenCode: `Agent Pipeline Orchestrator`.
- Works with files on the current Mac through a local node.
- Reads local files and folders when they are allowed.
- Creates user-facing output files, for example reports on Desktop or in the workspace.
- Runs checked local commands and simple test commands.
- Helps write and repair code in a sandbox/workspace.
- Sends heavier code-repair tasks into the configured server pipeline.
- Uses the heavy coder only inside the pipeline when the task requires it.
- Gives a clear final state: done, needs confirmation, denied, or failed.

## What It Does Not Promise Yet

- It is not a finished public product.
- It is not production-ready for unknown users.
- Linux and Windows support are not proven on real devices yet.
- GUI programs are not automated unless a real tool is connected.
- Jagannatha Hora is currently `not_connected`; AgentCode must not invent Lahiri/sign calculations.
- It should not be used unattended on sensitive projects, secrets, SSH keys, browser profiles, keychains, or system folders.
- It will not always understand vague tasks without clarification.

## The Core Rule

No evidence, no confident answer.

A good AgentCode answer must be based on something it actually checked:

- a file it read;
- a folder it listed;
- a command result;
- a created output path;
- a test/check result;
- or a clear tool denial.

If it cannot prove the result, it should say that it failed, needs confirmation, or needs one missing detail.

## Best First Use Cases

- "Read this file and summarize it."
- "Find this folder and make a tree/report."
- "Create a text report on my Desktop."
- "Write a small script and run it."
- "Check this project and fix one bug."
- "Run tests and tell me what failed."
- "Explain why this task cannot be done because a tool is not connected."

## Bad First Use Cases

- "Do everything in my real production project without supervision."
- "Use a GUI program that is not connected."
- "Work with secrets or private keys."
- "Make destructive changes without asking."
- "Guess where my files are if the path is unclear."

## How To Judge If It Worked

Good result:

- It used the right local tool.
- It created or changed the expected file.
- It tells the exact path of the result.
- It verifies the result after doing the work.
- It answers in Russian for Russian sessions.
- It admits when a tool is missing.

Bad result:

- It claims a file was created but gives no path.
- It says it read a file but did not.
- It searches the web for a local path.
- It guesses a folder location without checking.
- It repeats the same failed action.
- It gives calculations or app results without the required tool.

## Suggested Sales Position

Sell this as a private beta, not as a finished SaaS.

The honest offer:

"AgentCode is a local AI assistant for your Mac. It helps with files, folders, reports, scripts, and code tasks. I install it, connect it to server models, and tune it to your real workflows. It is useful now, but still beta, so we collect real tasks and improve it together."