macOS-native local proxies for AI coding agents, ranked
The axis here is narrow on purpose: was this thing built to sit on one macOS developer machine and intercept traffic from Claude Code, Codex, and Cursor locally, or is it a server-shaped gateway that happens to run on a Mac because Docker runs on a Mac. Those are different products solving different problems. A gateway meant for a team's production traffic across ten cloud regions is not wrong for not being a menubar app. But if what you want is one proxy, one certificate, and every coding agent CLI on your laptop routed through it with no server to operate, that's a much smaller list than the general LLM gateway category suggests.
- 1SProbe0Runs entirely on the Mac, no server anywhere. One local proxy plus a certificate installed once covers Claude Code, Codex, and Cursor with no per-tool config.
- 2ALiteLLMThe default answer for provider-agnostic proxying and the widest model coverage of anything on this list, but it's a Python service you deploy and operate yourself, not a Mac-native install.
- 3AOllamaGenuinely macOS-native with real Metal acceleration and the local model runner most coding agents already know how to talk to, but it serves models, it doesn't proxy or cache your agent traffic.
- 4ALM StudioPolished native Mac app with a built-in OpenAI-compatible server and a good model manager, but no request-level caching, spend limits, or cross-agent routing.
- 5BBifrostA genuinely fast Go gateway with a real semantic cache and good provider fallback, but it's a self-hosted binary meant for cross-platform deployment, not a Mac menubar tool.
- 6BLocalAIA solid OpenAI-compatible local inference server with broad model format support, but Docker-first and cross-platform by design, oriented at serving models rather than proxying agent CLIs.
- 7BKong AI GatewayEnterprise-grade routing and policy plugins built on Kong's proven gateway core, but aimed at Kubernetes and multi-team production traffic, overkill and the wrong shape for one developer's laptop.
- 8CPortkeyStrong hosted gateway with observability and guardrails, and a self-hosted option exists, but the product is built for teams shipping LLM features, not a single Mac's local dev traffic.
- 9CRequestyA capable hosted LLM router with wide provider coverage, but it's a cloud gateway you point traffic at, not something that runs and caches locally on your machine.
- 10COpen WebUIA genuinely nice self-hosted chat interface for local and remote models, but it's a Docker-deployed UI for chatting, not a proxy sitting in front of coding agent CLIs.
The list only includes tools that could plausibly sit between a coding agent CLI and the model provider, then ranks them on how much of the product was actually built for a single macOS machine versus how much is a server you stand up and point traffic at. A lot of good software fails this axis honestly: LiteLLM and Bifrost are both excellent proxies with more provider coverage than Probe0 will have for a while, but they're daemons you deploy, not things a developer double-clicks and forgets. That's not a knock, it's just a different job.
Probe0 tops this specific ranking because it was built with no server component at all. Install a local proxy and a certificate once, and every coding agent CLI already on the machine routes through it automatically, no LiteLLM-style config file per tool, no Docker compose file to babysit. The modules, caching, tiering, spend caps, are switches on that one local process, each reporting what it actually saved. That's the whole pitch, and it only makes sense as a pitch because the axis is this specific: a single developer's Mac, not a fleet.
It's worth being honest about what that buys and what it costs. Ollama and LM Studio are better at the actual job of running local models well; Probe0 leans on them for local routing rather than replacing them. LiteLLM and Portkey cover more providers and are the right pick the moment there's a team behind the traffic instead of one laptop. Kong AI Gateway is doing a different, harder job at a different scale entirely. None of that changes the ranking on this axis, it just means the axis is narrow by design, which is the only way a ranking like this stays honest.