feat(ai-providers): add llama.cpp and MLX as local providers

[datlechin]

Closes #1777.

What

Adds llama.cpp and Apple MLX as local AI providers, alongside the existing Ollama and custom OpenAI-compatible options.

Both llama-server (llama.cpp) and mlx_lm.server (Apple's mlx-lm) expose OpenAI-compatible HTTP servers (/v1/chat/completions, /v1/models, SSE streaming, no auth by default). TablePro's OpenAICompatibleProvider already drives exactly that contract, so this is an additive registry change, not new transport code. Ollama is actually the special-cased provider (native /api/chat, /api/tags, newline-JSON streaming); llama.cpp and MLX are more purely OpenAI-compatible and ride the standard path.

Changes

• AIModels.swift: new AIProviderType cases .llamaCpp and .mlx (display names llama.cpp / MLX, default endpoint http://localhost:8080, no API key, distinct SF Symbols).
• AIProviderRegistration.swift: register both in the OpenAI-compatible loop, so they appear in the Add Provider menu with endpoint config and model fetching.
• AISettingsView.swift: generalize the "Local" status label from Ollama-only to any local provider with no auth.
• Docs and CHANGELOG updated. New LocalProviderRegistrationTests mirroring the xAI registration test.

Design notes

• The generic custom provider remains the escape hatch for other OpenAI-compatible servers (vLLM, LM Studio, LocalAI) or remapped ports. Both llama.cpp and MLX default to port 8080, so the preset endpoint is a starting suggestion the user can edit.
• No auto-detection: Ollama's OllamaDetector is currently dead code (no call site), so parity is presets only.
• No native in-process MLX (mlx-swift) engine. That is a separate, large effort and not required to support MLX as a provider.
• For llama.cpp tool calling, the server must be started with --jinja (documented).

Test plan

• swiftlint --strict clean on changed files.
• New unit tests cover the enum arms, registry descriptors, OpenAICompatibleProvider construction, and the default endpoint resolving to /v1/chat/completions.
• Manual: start llama-server / mlx_lm.server, add the provider, Test Connection, fetch models, chat.

https://claude.ai/code/session_01FYwKEwKydeMG6WAHucMdY9

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
TablePro	🟡 Building	–	Jul 21, 2026, 4:42 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 76b655bf6d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Expose connection testing for the unauthenticated presets

When a user adds either new preset, this .none auth style makes AIProviderDetailSheet.authSection render EmptyView; the only existing Test Connection controls are inside the API-key/Cursor/xAI auth sections. Thus llama.cpp and MLX have no way to run the connection test that the new provider instructions explicitly tell users to perform, even after choosing a model. Add an unauthenticated test control (for example in the connection section) or remove that workflow from the provider documentation.

Useful? React with 👍 / 👎.