DATA · PERMISSIONS · TRUST

Your project stays on your computer. You grant the permissions. AI can work only when it receives the relevant data it needs.

This page explains the boundary between local data, ChatCode, the operational serverthe AI provider so you know exactly what you are allowing.

YOUR COMPUTERProject / Data
CHATCODEScope · Brain · Tools
AI PROVIDERRelevant context
Not: “upload the entire drive.” Instead: AI asks → ChatCode retrieves what is permitted → returns the context/result needed for the task.
1

Your repository is not bulk-uploaded to ChatCode servers for storage.

The project remains on the computer running ChatCode.

Project / files

They remain in your computer's filesystem. ChatCode works only with projects you added.

Project Brain

A local indexing/context layer that helps navigate the workspace efficiently.

ChatCode server

Used for licensing, accounts, updates and necessary connection/routing functions; it is not a repository storage service.

2

“Local” does not mean file content can never leave the computer.

When AI needs content for analysis, the context/results returned by a tool may be sent to the AI provider you are chatting with.

Important distinction: ChatCode avoids blindly uploading an entire project. But when you ask AI to read a file, debug code or summarize data, information needed for that task may become input to your AI provider. That provider's storage/training policies still apply.

For sensitive data, choose an appropriate provider/plan, review its policies, scope projects and permissions, and expose only data that is genuinely needed.

3

Permissions are real boundaries, not just prompt instructions.

Each project has an independent scope and permission set.

PermissionUse whenSafer default
Read/SearchUnderstand source/data, reviewAllow when needed
Edit filesFix/refactor/contentEnable only on target projects
Create/delete/renameCreate output, scaffoldOff for review-only work
TaskTest/build/lintAllow-listed tasks only
Git writeLocal stage/commitDoes not mean auto-push
Remote/SSHDeploy/operationsSeparate and least-privilege
4

Secrets and credentials must be treated differently from ordinary files.

ChatCode blocks many common sensitive paths and does not return remote passwords/private keys to the AI.

Local sensitive paths

Names/formats such as .env, .ssh, private key, credential store… are blocked by policy.

Remote credentials

AI selects a profile and allowed operation; ChatCode uses the password/private key internally without exposing it back to the model.

No filter replaces good permission design. If a deploy user can write across an entire server, legitimate operations by that user still have a large blast radius.
5

Safest pattern: Inspect → Plan → Change → Verify → Review.

Do not grant every permission and ask AI to “do everything” in the first message.

Inspect

Read-only. Confirm the project, state and root cause.

Plan

Ask AI to list files to change, tests to run and risks.

Change + Verify

Enable only necessary permissions, make minimal changes, verify, then review the diff.

Dùng ChatCode với project shop. Chỉ đọc và xác định root cause lỗi checkout. Liệt kê file dự kiến phải sửa và test cần chạy. Chưa sửa file.Tiến hành thay đổi tối thiểu theo kế hoạch. Chạy test liên quan. Trả diff và rủi ro còn lại. Không deploy.