Two security-first local AI tools for macOS in Swift 6

Related Projects
1

Hi Swift community,

I’ve published two small MIT-licensed Swift 6 projects for private, local AI workflows on macOS:

  • Ollama Local Client is restricted to literal 127.0.0.1:11434, fixed API paths, and a fixed model. It blocks redirects, proxies, cookies, cache, model substitution, malformed NDJSON, oversized responses, and cross-request cancellation.
  • Local Workspace Orchestrator performs read-only source analysis. File access is descriptor-relative with openat and O_NOFOLLOW; traversal, symlinks, sensitive paths, binaries, and oversized files fail closed. Model output has no approval or execution authority.

Both packages support macOS 14+, build under strict Swift 6, and have GitHub Actions coverage. The client has 10 security-focused tests; the orchestrator has 11 tests, including 500 concurrent reads under Thread Sanitizer.

I’d especially value feedback on the actor boundaries, URLSession cancellation design, and file-descriptor ownership model. The repositories include security policies and contribution guidance.