Security

Safeguards

These are things the current desktop app already does. The full wording is in Security.

Intro film · You approve each step: Tools and approvals, Computer use, Privacy and data, and Account and profile. Every frame is a real screenshot of Koder 0.4.45. The push-ins and dissolves were added in post. The film does not show AI actually changing a file or using the computer. There is no voiceover. The music was written by a program.
Settings, Tools and approvals: careful by default, confirm every file change, and ask before every commandSettings, Tools and approvals: careful by default, confirm every file change, and ask before every command Settings, Computer use: the switch that lets AI use this computer is off, and every action asks firstSettings, Computer use: the switch that lets AI use this computer is off, and every action asks first

1 / 2

Screenshots: Settings → Tools and approvals: before AI acts, the default is Careful. File changes confirm every time and are written only after you accept. Commands ask every time. Computer use is off. After you turn it on, every action asks.

Confirm before it changes anything

Before it writes or edits a file, it shows the diff. It writes only when you click Accept. A command shows the full text and asks, by default. Deletes, privilege changes, disk operations, shutdown and restart, force-push and rewriting git history, and .env and key files always ask, no matter the setting. A confirmation waits at most 30 minutes. After that it counts as a no.

It listens only on this computer

The interface listens only on 127.0.0.1, and it refuses requests that come from other websites. The kernel generates a new token every time it starts. The token file can be read only by the current system user. Without the token, nothing can call the kernel.

Keys stay on this computer

Model keys are stored with AES-256-GCM. The key file used for that encryption can be read only by the current user. The interface does not show any part of a key. Koder does not read model keys from environment variables or from a .env file. It is not in the macOS Keychain or Windows Credential Manager yet.

The work has a boundary

Each task works only in the folder you name. You can switch it to Ask or Plan, which are read-only. You can also turn a tool off on the Tools page. That takes effect immediately. KARMEN bot only chats. It does not use tools.

Computer use is off by default

It can screenshot, click, and type only after you turn it on. Every action asks by default. There is a Stop button at the top of the screen, and ⌃⌥⌘. stops it immediately. Screenshots go only to the model service you chose.

Web pages do not reach your local network

When it fetches a page, only public http and https are allowed. Local and private-network addresses are refused. A model endpoint cannot include a username or a password.

The installer itself is not code-signed. That is a separate point, on install safety. Approvals are not loosened because the installer is unsigned.

What each of the three modes lets through

Careful: file changes ask, and commands ask. Balanced: writes inside the task folder go straight to disk, and commands still ask. Automatic: both run on their own. None of the three automatically lets through deletes, privilege changes, disk operations, shutdown and restart, force-push, rewriting history, handing text to a shell, or .env and key files.

If a model asks to switch work modes, it asks you first. If you do not answer a choice card in 60 seconds, it continues on its own judgment, and you can still add an answer later. These rules are written on this computer. The public API does not change them.