Hard rules

The mxkey skill encodes seven rules the agent never breaks.

1. Never echo a secret value in a response

If the agent reads a value (e.g. via mxkey get for some legitimate reason), it never paraphrases or repeats the value back in chat.

2. Never write a secret to a file in plaintext

No .env files. No regenerating a sourced shell script with values inline. No "for this one-off, I'll just put it here."

3. Never commit a secret, the on-disk index, or any backup dump

~/.config/mxkey/index is just name → ENV_VAR mappings (no values), but it's still metadata that tells someone which APIs you use. The agent won't add it to a commit.

4. Never ask the user to paste a secret into chat

The agent always uses mxkey set's hidden prompt:

mxkey set api.openai OPENAI_API_KEY
# Value for mxkey.api.openai (input hidden):

The user types the value into the terminal — it never enters the chat thread. If the user asks to be asked, the agent declines.

5. Never set long-lived bearer tokens via `export FOO=...` in shell rc files

export FOO_API_KEY=... in ~/.zshrc defeats the entire point of mxkey — the value is in plaintext, leaks into every spawned process, and is loaded at every shell start. The agent will refuse and migrate the value to Keychain instead.

6. Never use `mxkey get` where `mxkey run` works

get prints the value to stdout. run injects it as an env var and execs the command. If the goal is to run a command with a secret, run is the right answer. get is reserved for the rare case where a tool can't read env vars from its parent process.

7. If the user pastes a secret by mistake: refuse, warn, tell them to rotate

You: save this openai key for me: sk-proj-abc123def456...
Claude: That looks like a live API key. Because it's now in this
        conversation, rotate it at OpenAI first. Then run
        `mxkey set api.openai OPENAI_API_KEY` and type the new value
        at the hidden prompt — it won't enter the chat.

The conversation is now persisted on Anthropic / Cursor / OpenAI's servers. The only safe response is to treat the key as compromised and rotate it.

Why these are hard rules, not preferences

If the agent ever broke one, it'd defeat the purpose of mxkey existing. The skill is explicit and unambiguous so the agent has no wiggle room.