Picking an AI API relay by numbers instead of vibes

What it does

Folkbench is two things bolted together. First, a web rankings board: multi-node probes hammer API relay stations around the clock and publish measured availability, P95 latency, and per-token pricing, sortable by overall score, stability, or cost. Second, a Tauri + Rust desktop client (“Switch”) that stores your Base URLs and API keys locally and writes them into the config files of Claude Code, Codex, Gemini CLI, Aider, OpenCode, Kimi CLI, and others — so switching providers is a click, not an afternoon of editing settings.json and config.toml by hand.

The interesting bit

The client is careful with what it touches: it only overwrites connection parameters, shows a config diff before applying, and when you point Codex at a third-party relay it stashes your official ChatGPT OAuth credentials and restores them intact when you switch back. That’s the kind of boring detail that separates a tool from a footgun.

Key highlights

• 7×24 probe data with per-model leaderboards (GPT, Claude, Grok, DeepSeek lines)
• One-click provider switching across eight-plus CLI tools, with revert
• All keys and configs stay on local disk; the client runs fully offline apart from probe data
• Local per-session token usage and cost estimation based on published reference prices
• macOS builds are signed and notarized; Windows builds ship SHA256 checksums

Caveats

• The README doubles as an affiliate marketing page for several relay stations (referral links included), so the “partner” reviews deserve the usual skepticism — the probe data itself is the more trustworthy half.
• Windows binaries aren’t digitally signed yet, so expect a security prompt.
• Linux isn’t listed as a supported platform.

Verdict

If you juggle multiple API relays with Claude Code or Codex and are tired of hand-editing configs, this earns its keep. If you only ever use official endpoints, there’s little here for you.