Every AI coding tool wants its own config. Devin reads ~/.config/devin, Claude reads ~/.claude, Cursor keeps ~/.cursor/mcp.json, Codex wants TOML, Continue wants YAML. Set up the same MCP server in five tools and you maintain five copies of it - and five places for an API key to leak.

universal-ai-config is our answer: one unified store at ~/.agents/ that every provider reads from - or that you explicitly export to each provider's native format.

What it does

The security story

v0.2.1 shipped a fix for a HIGH finding (CWE-427, uncontrolled search path): project_config() used to accept any .ai/ directory it walked into. In a shared or world-writable directory, another user could plant a config.json that merged into your settings - injecting provider credentials and arbitrary MCP servers, which sync would then write to provider files.

Now .ai/ is trusted only when both the directory and its parent are owned by the current user; anything else is skipped with a SecurityWarning.

Try it

uv tool install universal-ai-config   # or pipx / pip
ai-config migrate   # pull existing configs into ~/.agents/
ai-config status