I'm still mulling over the right way to add authentication support to this, while managing credentials securely.
One possibility is to have a set of flags --claude/... for each AI agent, which turns the CLI parameters from "commands or URLs" to "name of the MCP registered with that AI tool". At which point we can grab the invocation settings from one Claude file and also dig the credentials out of another. It's not ideal and is poking at internals a little, but would avoid us needing credentials store access or encouraging use of an AI agent anywhere near where it can touch regular credentials.
Proposals for better support welcome.
My main requirements are:
- must be suitable for running "headless"; I do AI work via "ssh someai@localhost" to isolate its access while still letting it see content I choose to share
- must not add a huge dependency chain
- not platform-specific, or with sane platform fallbacks
- I use Linux laptops; the macOS Keychain is fit for purpose, the DBus Keyring is not fit for any of my purposes
I'm still mulling over the right way to add authentication support to this, while managing credentials securely.
One possibility is to have a set of flags
--claude/... for each AI agent, which turns the CLI parameters from "commands or URLs" to "name of the MCP registered with that AI tool". At which point we can grab the invocation settings from one Claude file and also dig the credentials out of another. It's not ideal and is poking at internals a little, but would avoid us needing credentials store access or encouraging use of an AI agent anywhere near where it can touch regular credentials.Proposals for better support welcome.
My main requirements are: