Decide whether you can maintain the Cursor connection
The repository was archived on October 20, 2026 and is read-only. That changes the adoption decision: choose cursor-byok only if retaining Cursor's Agent interface justifies evaluating and potentially maintaining the gateway yourself. The supplied v1.0.1 release page offers platform installers, but an available installer does not establish compatibility with a later Cursor release.
The README describes a local service that forwards Cursor requests to configured model APIs and adapts OpenAI- and Anthropic-compatible protocols. It advertises tool calling, Skills, MCP, and multi-turn conversations. Those are project claims; our sandbox did not verify them in a live editor session.
What our September sandbox established
Our run on September 27, 2026 used commit 2068ab2, whose full hash is 2068ab20513288e6a0289d33febe872a30608610. The sandbox record identifies lab-rust:1, 3 CPUs, and 12 GB of memory. Installation completed in 28 seconds, and compilation completed in 381 seconds. Our test command ran for 143 seconds and exited with status 101.
Our harness stores each command's outcome, exit code, elapsed time, timeout status, and output tail, alongside parsed test summaries. The attached stored parser summaries report 368 passed, 2 failed, and 370 total. The retained output repeats summaries after a replay marker. These aggregates may double-count suites; they do not establish unique or exhaustive test counts, and we have not replaced them with a sum of the visible footers.
The raw log separately identifies trace_producers_do_not_wait_for_sqlite_and_artifacts_stay_ordered in cursor_trace_queue as a failing test. It panicked at server/tests/cursor_trace_queue.rs:90:6 after Result::unwrap() received Elapsed(()). This is more specific than saying the assertion was unavailable, but it still does not tell us whether scheduling, the environment, or application behavior caused the timeout.
The record contains no live Cursor acceptance test, desktop installation check, provider-cost benchmark, or security audit. Audit coverage is unknown. A successful build does not fill those gaps.
The MrKeyoor testing methodology explains command recording and the limits of sandbox results. That page provides provenance context; its current procedure is not evidence that this older run used today's methodology version. No new execution was performed for this correction.
Configuration checks and Agent routing are separate
The README's setup path is to download an installer, enter an endpoint, key, and model ID, test the configuration, start the service, and restart Cursor into a new conversation. Model configurations can also specify context limits, maximum output tokens, reasoning effort, custom headers, and extra request parameters.
Issue 465 documents a Windows user's successful built-in model check followed by Cursor's "The model you chose is not available" error. The reporter recorded model-test calls but no real Agent traces. This is a user report about cursor-byok 1.0.0, not a failure reproduced in our sandbox.
The same issue includes a September 20 comment from the reporter saying the problem was resolved after merging 433. It should therefore inform an acceptance check, rather than serve as proof that the problem remains unresolved in v1.0.1. The supplied evidence does not independently confirm the fix across other environments.
The report also shows a local CA and Cursor proxy settings, including disabled strict SSL verification. Check the required changes against your workstation policy before installation. These details describe the reported Windows setup; our Linux build did not test certificate installation or restoration of editor settings.
Local credentials still lead to provider requests
The README says API keys and application settings are stored locally, while requests still go to the configured model provider. Local storage does not make a remote provider into local inference. Review that provider's handling of the prompts, code context, and tool results you intend to send.
The gateway is independent of Cursor and is not endorsed by its developers. Its README also distinguishes the free software from model APIs that may charge for usage. We measured neither savings nor provider billing accuracy.
Choose a replacement client or a separate gateway
Cline supports provider choice within its own coding-agent clients. Its README lists a VS Code extension with tools and human approval, plus terminal, desktop, and JetBrains options. Consider Cline if you can change the agent interface instead of preserving Cursor's Agent workflow. Its provider list is documentation, not a compatibility result from our run.
Continue is available as a CLI, VS Code extension, and JetBrains plugin, but its supplied repository notice says it is no longer actively maintained and is read-only. It is an option for evaluating or maintaining an existing codebase, not an answer to the requirement for active upstream maintenance.
LiteLLM provides a Python SDK and a centralized model API gateway. Consider LiteLLM when several applications need shared provider access. It addresses a different layer and could complement a client integration; the supplied source does not establish that it replaces cursor-byok's Cursor-specific routing.
What to check before relying on it
Use a disposable workspace and your intended Cursor version. Confirm that a real Agent request reaches the gateway, then exercise file tools, the MCP connections you require, conversation continuation, and subagents if your workflow uses them. Restart the editor and repeat the checks. These are recommended checks, not tests we performed.
Keep the installer and configuration for a combination you have verified, and document how to restore the editor's network settings. With an archived upstream, decide who will investigate the next incompatible client change before depending on the gateway.
How to read the retained ratings
The stored ratings below are historical editorial assessments on a five-point scale: 1 means substantial barriers, 2 means major limitations, 3 means usable with material caveats, 4 means strong evidence within the assessed dimension, and 5 means unusually complete evidence. Setup concerns installation and configuration; documentation concerns the available instructions; community concerns maintenance and support activity; maturity concerns demonstrated reliability. These are editorial judgments, not benchmark results or adoption measurements.
The retained community rating of 4 cannot describe present maintenance: the archive notice supersedes the earlier activity narrative. The supplied extracts also do not verify the earlier claims about the linked guide's language or a missing contributing guide. Read those retained scores with these evidence limits. Our failed test command and absence of a live editor check do not support a claim of broad compatibility.