Limitations

OSW is a local LLM proxy with clear boundaries. The following are deliberately not done, not left unfinished.

Network & access

No system-level proxy: OSW does not take over global traffic; it serves only clients that explicitly point their Base URL at it.
No remote access: it is designed to run on this machine, listening on 127.0.0.1 by default. Exposing it to a LAN is a risk you take on yourself (it carries no auth).
The management service listens locally only: the management API always binds to 127.0.0.1; to access it from another machine, use an SSH tunnel.

Gemini's /v1beta/models/* is not supported: connect via the OpenAI-compatible or Anthropic shape instead.
Protocol conversion is a compatibility layer: no conversion by default. Only after you enable it explicitly may an endpoint accept requests in other protocols, and some parameters may be dropped.

Failover happens only within a group: the models attached to one logical model, or one channel with conversion explicitly enabled. It does not fall back across groups — that is routing's job.
No stitching after streaming starts: a mid-stream break aborts; the output of multiple channels is never stitched together.

No team collaboration / multi-user permissions: this is a single-machine tool.
No account system: no login; cloud sync uses your own credentials.
Cloud sync only encodes, never encrypts: base64 keeps the file from being readable at a glance, but is not a means of secrecy (see Config cloud sync).
No automatic / scheduled sync: upload and pull are one-shot actions you trigger manually, and there is no conflict merging.

Only non-streaming JSON: streaming (SSE) and non-JSON content do not participate in rewriting (see Request rewriting).
The response stage is currently disabled: because a streaming response cannot be rewritten at the event level and it accounts for the vast majority of traffic, the response stage is disabled entirely — the editor doesn't offer it, and saving or dry-running a rule with a response stage is rejected (see Request rewriting).

Bodies are recorded verbatim: capture is not size-limited, and the request body is read fully into memory; request bodies are also not masked (see Request logs).

No auto-update for major versions: versions with breaking changes (data and config no longer carried forward) must be downloaded and installed manually.
No auto-install on macOS: because of ad-hoc signing, it can only check for updates and open the DMG download page.

If your need falls outside these boundaries, feel free to open an issue in the GitHub repository to discuss it.