ccswitch handles long-lived OAuth credentials. This doc captures the threat model and the residual risks the shell implementation cannot fully eliminate.
ccswitch protects credentials from:
- Other users on the same machine — credentials live in the user's login Keychain (or 0600 files), inaccessible to other UIDs.
- Cloud breach — when the 1Password Connect backend is enabled, the local daemon talks to a Connect server through Cloudflare Access. The bearer token + CF Access service-token secrets are stored in the login Keychain.
- Process accident — config TOML never contains secrets; only references to keychain entries. A leaked config file does not expose credentials.
ccswitch does not defend against:
- A process running as the same UID as the user. Such a process can already
call
security find-generic-password, read~/.claude/.credentials.json, attach via ptrace, or read environment variables of any other process the user owns. Argv visibility (next section) is strictly weaker than this.
Two code paths briefly pass secrets through subprocess argv on macOS, where
ps -ef reveals full argv to same-UID observers. We mitigate where we can;
where we cannot, we accept the residual risk.
Closed. Curl gets its bearer + CF Access headers via a config file fed
through process substitution (-K <(...)), and request bodies arrive on
stdin (--data-binary @-). Neither tokens nor credential JSON appear in
the curl process arg list. Verify with:
ps -ef | grep curl
# headers and body are absent; only the URL and HTTP method are visibleOpen, by Apple design. The macOS security(1) CLI accepts the password
only via -w VALUE (in argv). The interactive bare -w form requires a
double-prompt with retype confirmation that is unusable from non-interactive
scripts. There is no documented stdin variant.
The credential value therefore appears in the security subprocess's argv
during a single short-lived invocation (~50 ms). Same-UID ps polling
could observe it; cross-UID observers cannot.
To eliminate this entirely, ccswitch would have to call the Security
framework directly (e.g., via a small Swift helper, the Go keyring
package, or linking libSecurity.dylib). A future native rewrite will
do so; see the Go-rewrite proposal in the repo root.