Can Claude Code actually encrypt and decrypt my files?

Yes. The bge command-line tool ships a Claude Code skill (skills/bge-encryption). Once the CLI and the skill are installed, you ask in plain language — for example, 'encrypt report.pdf for Alice' or 'decrypt invoice.bge and summarise it' — and the agent runs bge to do it, locally, in the open .bge format.

Is it safe to let an AI agent handle passphrases and keys?

The skill is written to be safe in an automated shell. It never puts a passphrase on the command line or in shell history (it pipes it via --password-stdin), never overwrites files or generates keys without confirming, and prefers decrypting to a pipe so plaintext is not left on disk. The bge address book holds public keys only — it never stores a private key — so decrypting still requires a private-key file you provide.

Does it work with ChatGPT or AI agents other than Claude Code?

Yes. The packaged skill is for Claude Code, but bge itself is a plain, scriptable command-line tool: deterministic flags, stdin/stdout piping, --password-stdin, and standard sysexits exit codes. Any assistant that can run shell commands can drive it the same way — ChatGPT's coding/agent tools (such as Codex), editor agents, or a CI script. You can also hand any model the skill file as a plain instruction set.

Can the agent see my private key?

The CLI uses file-based keys only and never stores private keys. To decrypt an identity-locked file, you point bge at a private-key file you control; the agent does not have a key store to read from. Encrypting by name uses the public-key address book in ~/.bge, which holds no private keys.

What do I need installed to use the bge skill?

Two things: the bge CLI on your PATH (install with brew install --cask dotbge/tap/bge on macOS, or download a build for macOS/Linux), and the skill copied into ~/.claude/skills/ (or a project's .claude/skills/). The skill checks that bge is installed and tells you how to install it if not.

bge for Claude Code, ChatGPT & AI agents

Strong file encryption usually means remembering flags and juggling key files. The bge-encryption skill teaches Claude Code how to drive the bge command-line tool — which mode to use, how to look up saved contacts, how to feed a passphrase without leaking it, and what each exit code means. You describe the outcome; the agent runs the right command, on your machine, in the open .bge format. The same commands work from ChatGPT's coding/agent tools or any other assistant that can run a shell — see below.

What you can ask

Plain-language requests, and the command the agent runs for each:

💬

“Encrypt the contract for Alice.”

Finds Alice in your saved contacts and seals the file to her public key — only she can open it.

bge encrypt contract.pdf -r Alice          # → contract.pdf.bge

💬

“Encrypt my tax docs so only I can open them.”

With no recipient, bge seals to your own active identity — a one-key backup only your private key unlocks.

bge encrypt taxes-2025.pdf                  # → sealed to you

💬

“Password-protect notes.txt with the passphrase in $PW.”

Pipes the passphrase in via --password-stdin — it never lands on the command line or in your shell history.

printf %s "$PW" | bge encrypt notes.txt --password-stdin

💬

“Decrypt invoice.bge and pull out the total.”

Decrypts to a pipe, not a file — the plaintext is read by the next tool and never written to disk.

bge decrypt invoice.bge -k me.pem -o - | grep -i total

💬

“Who is secret.bge encrypted for?”

Reads only the public header — no key needed — and names the recipient if they're in your address book.

bge inspect secret.bge                      # mode · key ID · Encrypted for: …

💬

“Encrypt this whole folder for Bob, as one file.”

Zips the folder and encrypts it to a single archive that also hides the file names.

bge encrypt project/ -a -r Bob              # → project.zip.bge

💬

“Encrypt this for Carol.” (Carol isn't saved yet)

Doesn't guess. It asks you for Carol's public key or .bgekey card, saves the contact, then encrypts.

bge contact add carol.bgekey               # then: bge encrypt … -r Carol

The guardrails it follows

The skill is written for a shell with no human watching, so it builds in the safe defaults you'd want:

●No silent hangs. It never uses a bare interactive -p prompt that would block forever — password mode always goes through --password-stdin.
●Secrets stay out of sight. Passphrases are piped from a variable or file, never written as a command-line argument (which would leak into the process list and history).
●It asks first. Before overwriting a file (--force), generating new keys, or writing decrypted plaintext to disk, it checks with you.
●Public keys only. The address book in ~/.bge never holds a private key, so the agent can encrypt by name but can't decrypt without a key file you hand it.
●No stray plaintext. When the goal is to read or transform a file, it decrypts to a pipe (-o -) instead of leaving the cleartext on disk.

Install the skill

First, make sure the bge CLI is on your PATH:

macOS · Homebrew

brew install --cask dotbge/tap/bge

Then add the skill — it lives in the bge-cli repository under skills/bge-encryption:

# Personal — available in all your projects:
mkdir -p ~/.claude/skills && cp -r skills/bge-encryption ~/.claude/skills/

# …or per-project, checked into a repo:
mkdir -p .claude/skills && cp -r skills/bge-encryption .claude/skills/

That's it. Ask Claude Code “encrypt report.pdf for Alice” and the skill loads on its own. If bge isn't installed yet, it tells you how and stops rather than guessing.

Claude Code, ChatGPT, or any agent

The packaged skill is built for Claude Code, but bge itself is a plain, scriptable tool — deterministic flags, stdin/stdout piping, --password-stdin, and standard sysexits exit codes. Any assistant that can run a shell command drives it the same way:

Claude Code — drop in the skill (above) and just ask; it loads automatically.
ChatGPT — its coding / agent tools (such as Codex) can run the same bge commands, or hand ChatGPT the skill file — it doubles as a plain instruction set any model can follow.
Other agents & CI — editor agents, automations, or a build script can call bge non-interactively the same way.

See the CLI reference for the full command set.

View the skill on GitHub CLI reference →