| [ Web Proxy ] |
| Viewing: https://commandcode.ai/docs/resources/security | [Back] [Original] |
Command Code is designed with security and privacy as core principles. Your code stays yours - we never train on it, and you control exactly what Command Code can access.
| What | Stored? | Where | Used for training? |
|---|---|---|---|
| Source code | Never stored | Your machine only | No |
| Taste profile | Local + optional cloud sync | .commandcode/taste/ and commandcode.ai | No |
| Conversation history | Local only | ~/.commandcode/projects/ | No |
| Authentication tokens | Local only | ~/.commandcode/auth.json | No |
| AGENTS.md | Local only | Project root | No |
Command Code does not train on your code. Taste learning runs locally and stores preferences as structured rules - not code snippets.
Command Code uses a permission system that puts you in control of every action. By default, any action that modifies your system requires explicit approval.
Command Code has five permission modes.
| Mode | File reads | File writes | Shell commands | When to use |
|---|---|---|---|---|
| Default | Allowed | Requires approval | Requires approval | Day-to-day work |
| Auto-Accept | Allowed | Allowed | Safe filesystem commands only (mkdir, cp, non-recursive rm, ); other shell still asks | Trusted iteration |
| Plan | Allowed | Blocked, except a plan file under ~/.commandcode/plans/ | Read-only only (e.g. git status) | Exploring and designing |
| Bypass | Allowed | Allowed | Allowed | Throwaway environments only; deny rules and the root/home removal check still apply |
| Dont-Ask | Allowed | Only pre-approved | Only pre-approved | Unattended runs where a wrong prompt should fail closed instead of waiting |
Switch modes with shift+tab inside a session, or start with a specific mode:
cmd --plan # Start in plan mode
cmd --auto-accept # Start in auto-accept mode
cmd --permission-mode auto-accept # Start in auto-accept mode (long form)
cmd --permission-mode dont-ask # Start in dont-ask mode
cmd --yolo # Start in bypass mode (use with care)
Deny rules (and explicit ask rules) always win, in every mode, including bypass. See Permissions for the full decision pipeline, rule syntax, and the complete mode decision table.
The first time you run cmd in a project, Command Code asks whether to trust that directory. This prevents accidental execution in untrusted locations.
cmd --trust # Skip the trust prompt
When running in headless mode (cmd -p), all write operations are blocked by default. This keeps CI/CD pipelines and scripts safe.
To enable writes in headless mode, you must explicitly pass:
cmd -p "fix lint errors" --yolo
--dangerously-skip-permissionsis also accepted as an alias.
Only use --yolo in trusted environments like your own CI pipelines. It allows Command Code to modify files and run shell commands without confirmation.
Command Code connects to the internet for:
cmd login)Command Code does not make any network requests for telemetry or tracking without your knowledge. See Telemetry for details on what optional telemetry is collected and how to opt out.
Command Code only accesses files within:
cmd)--add-dir or /add-dir~/.commandcode/It does not read files outside these boundaries.
Authentication credentials are stored locally at ~/.commandcode/auth.json. They are never sent to any third party.
When connecting MCP servers, be aware that:
cmd mcp addcmd mcp list or /mcpOnly connect MCP servers you trust.
Command Code creates checkpoints before every file modification. If anything goes wrong, you can instantly rewind to a previous state.
Esc twice to open the checkpoint selectorFor organizations with stricter requirements, Command Code Enterprise offers:
Contact support@commandcode.ai for details.
| Web Proxy Viewer | New URL | Original Page |