[Ganesh75-web]

Problem

Ask AI works only with ChatGPT, Claude Code and OpenRouter. Many people use a provider Folio doesn't list (Z.ai, DeepSeek, Mistral, Groq, a company gateway) or a model on their own computer (Ollama, LM Studio). All of them accept OpenAI-style Chat Completions, which Folio's OpenRouter client already sends, but the OpenRouter address was fixed.

Change

A Custom endpoint provider in Settings → Ask AI, with a base URL, an optional API key and a model ID. It reuses the OpenRouter client's Chat Completions code:

• Shared code: electron/chat-completions.cjs holds the stream reader, the error messages, the secure-storage check and the write ordering. Both openrouter.cjs and the new custom-endpoint.cjs use it. OpenRouter's behavior and its existing tests are unchanged. ModelPicker is shared by both settings screens.
• Connecting: Folio checks GET {base}/models before it saves anything.
  • The key is stored only with system credential protection, the same as the OpenRouter key.
  • Servers without a key get no Authorization header.
  • The model list comes from /models when the server has one; otherwise you type a model ID.
• Addresses (the first requirement from Ask AI: connect any OpenAI-compatible endpoint (base URL + key + model) #78):
  • The base URL must be https://. Plain http:// is accepted only when every address the name resolves to is on this computer or the local network. Its requests also run behind a local-only guard, so if the name later resolves to a public address (a company name off the VPN, for example), nothing is sent.
  • Any address that resolves there, http or https, goes through the existing "Open a local address?" confirmation (confirmLocalAddress in main.cjs, now passed to ai.register) before Folio sends anything. Declining sends nothing.
  • The approved origin and addresses are saved with the endpoint. Each request runs in its own in-memory session behind network.cjs's guardSession, the guard URL downloads use. So requests reach only the approved local addresses, and Folio connects to addresses it resolved itself.
  • If a name later resolves to an unapproved local address, the request is refused, and the message asks you to connect it again.
• Not allowed in the address: credentials, query strings and fragments. Redirects are refused.
• Ask AI panel: shows the model and "Sent to {host}". Passage questions send the passage instead of the whole document, the same as OpenRouter.
• README: setup and privacy sections updated.

Issue or approval

Approved in #78: #78 (comment) (Matthew Berman, "Approved as you scoped it"), with two requirements:

1. https except for local addresses, which go through the existing local-network confirmation. Done as described above.
2. CI on Windows, macOS and Linux. I tested on Windows only; please check macOS and Linux.

How I checked it

All on Windows 11 with Node 26.5. I didn't test on macOS or Linux.

• npx tsc --noEmit and npm run build pass.
• node --test scripts/test-custom-endpoint.cjs scripts/test-openrouter.cjs scripts/test-network-security.cjs scripts/test-claude-path.cjs: 27 passed, 0 failed (1 Unix-only test skipped). There are 8 new tests for the endpoint and 2 for the guard. They use a stand-in for the system's encryption. They cover:
  • http refused unless every address is local (a name with one local and one public address is refused), and refused later if the name moves to a public address;
  • the question asked before anything is sent, and declining sends nothing;
  • the approval saved and passed on every request, and a hand-edited http config without an approval isn't trusted;
  • the key encrypted, a rejected key keeping the working setup, and a failed save keeping the old key with the old address;
  • keyless servers;
  • optional model lists;
  • streaming, citations, passage questions, errors that name the server, and Stop;
  • a clear message when an answer runs past 5 minutes, and a 404 that points at the base URL (such as a missing /v1).
• Smoke check (scripts/smoke-custom-endpoint.mjs, added to smoke-test.mjs): runs the real app against a real HTTP server on folio-local.test (the existing host-resolver rule). It covers the question being asked and declined with nothing sent, then connecting, models, a streamed answer, a citation jumping to page 2, Stop closing the connection at the server, choosing a typed model ID, and removal. This passes locally.
  • The full npm run smoke stops later at the highlight step (smoke-test.mjs, .highlightEditor). It stops at the same step on unmodified main on this machine, so I think it's a local environment difference (Node 26 here, Node 22 in CI), not this change.
• Live, not committed: I drove the built app with no network mocks.
  • A Z.ai GLM Coding Plan (https://api.z.ai/api/coding/paas/v4, model glm-5.3-flash) answered "The ferry leaves at dawn (p. 2)" in about 2 seconds, and the citation jumped to page 2. No local-address question was asked.
  • A keyless server on http://127.0.0.1 triggered the question once, received no Authorization header, and streamed its answer.

Screenshots of the settings screen and an answer are attached below.

Code written by Claude Opus 5.5 in Claude Code. I reviewed and tested it.

🤖 Generated with Claude Code

[Ganesh75-web]

Screenshots from the Windows run (they're hosted in a gist, since the GitHub CLI can't attach images):

Settings → Ask AI → Custom endpoint, connected to a Z.ai endpoint with a model chosen:

Ask AI panel, with an answer from that endpoint and its page citation. The footer shows the model and "Sent to api.z.ai":

[Ganesh75-web]

Before screenshot, to go with the ones above (CONTRIBUTING asks for before and after). Same screen and window size, from unmodified main (0590fac) on Windows:

Before: Settings → Ask AI. The provider list has only ChatGPT, Claude and OpenRouter, and OpenRouter's address is fixed:

After: the "Custom endpoint" screenshot in the comment above.