Security

Security considerations when using prompts-gpt.

Credential Storage

Token Storage

Tokens are stored in .prompts-gpt/.credentials.json:

{
  "token": "pgpt_abc123",
  "apiUrl": "https://prompts-gpt.com"
}

File permissions set to 0600 on Unix (owner read/write only)
Auto-added to .gitignore on creation
Never committed to version control

Token Security

Tokens must start with pgpt_ prefix
The SDK refuses to send tokens over unencrypted HTTP (except localhost)
Tokens are validated before API calls

Permission Auto-Fix

If the credentials file has overly permissive permissions, the SDK automatically tightens them to 0600 on load.

Git Safety

Destructive Git Operations

By default, prompts-gpt blocks destructive git operations during execution:

{
  "disallowDestructiveGit": true
}

Blocked operations:

git stash
git reset --hard
git checkout --force
git clean -fd

Override with --allow-destructive-git flag.

Sandbox Modes (Codex)

Mode	Permissions
`workspace-read`	Read-only access to workspace
`workspace-write`	Read/write to workspace files
`full-auto`	Unrestricted access

prompts-gpt run task.md --provider codex --sandbox workspace-read

Permission Modes (Claude)

Mode	Permissions
`default`	Prompts for each action
`acceptEdits`	Auto-accepts file edits
`bypassPermissions`	Unrestricted

prompts-gpt run task.md --provider claude --permission-mode acceptEdits

Network Security

HTTPS Enforcement

The SDK enforces HTTPS for all API communication:

const client = new PromptsGptClient({
  apiUrl: "http://example.com",  // Rejected (not HTTPS)
  apiUrl: "http://localhost:3000",  // Allowed (localhost exception)
  apiUrl: "https://prompts-gpt.com",  // Allowed
});

Request Timeouts

All requests have configurable timeouts with a maximum cap:

Default: 30 seconds
Max: 600 seconds (10 minutes)
Generate endpoint: 60 seconds

Data Privacy

What is Sent to the API

Project token (for authentication)
Prompt query parameters (for sync/pull)
Generation input (goal, context) for generate command

What is NOT Sent

Source code
Agent output
Run artifacts
Local configuration
Environment variables

Offline Operation

Many commands work entirely offline:

run, run-batch, sweep, orchestrate — local execution only
doctor, status, validate — local checks
diff, list — local file inspection

Commands that require API access:

sync, pull — fetch prompts from API
generate — AI-powered prompt generation
init — token validation

CI/CD Security

Token as Secret

Always store tokens as CI secrets:

# GitHub Actions
env:
  PROMPTS_GPT_TOKEN: ${{ secrets.PROMPTS_GPT_TOKEN }}

Prefer PROMPTS_GPT_TOKEN in the remote environment over committing or baking .prompts-gpt/.credentials.json into the runner image.

Remote Readiness Audit

Before turning on a remote runner, validate the environment:

prompts-gpt doctor --remote

This checks:

Non-interactive execution readiness
Secret-based token presence
Safe API URL usage
Provider auth environment variables
Destructive git protection
Artifact directory gitignore coverage
Local credential file leakage risk

Minimal Permissions

Use the least permissive provider settings in CI:

prompts-gpt run task.md --provider codex --sandbox workspace-read