About SwanCat
One indie dev. Two tools. Zero token waste (well, −70% measured).
一位独立开发者,两个工具:网关帮你省 token,搜代码的帮你省时间。
Products / 产品
| What | Status | Link | |
|---|---|---|---|
| HiRouter | BYOK AI gateway: one endpoint for all models, smart routing + cache. Measured −70% token spend over 170 real requests. | Free public beta, 500 req/day | Try it · Docs |
| RLS | Local code search for AI agents: one exe, CLI for humans + MCP for AI. | v0.1.0, free | Docs |
| 是什么 | 状态 | 入口 | |
|---|---|---|---|
| HiRouter | BYOK AI 网关:一个地址接所有模型,智能路由 + 缓存,170 次真实请求实测省 70%。 | 免费公测,500 次/天 | 试用 · 文档 |
| RLS | 给 AI Agent 的本地代码搜索:单个 exe,人用 CLI,AI 走 MCP。 | v0.1.0,免费 | 文档 |
Why HiRouter / 为什么用 HiRouter
BYOK, bills stay yours — your upstream keys, your upstream bill. The gateway never resells tokens, so its incentive is aligned with yours: the fewer tokens sent, the more you trust it.
Measured −70.8% — 170 real Codex requests: 23.18M tokens in, 6.77M sent upstream (client-side estimate, not billing; method and caveats).
One endpoint for all models — routing + automatic failover + connection reuse behind a single gateway key.
Saves where it hurts most — fold + summarize + rewrite at the gateway; 132 of 170 requests saved 70%–90%, larger contexts save more.
Free public beta — 500 requests/day, 60/minute, no credit card.
Private by design — tenant-isolated, keys/passwords/sessions stored as hashes, no prompt bodies on disk, rotate or revoke keys anytime.
3-minute setup — CC Switch custom provider (
https://api-hirouter.swancat.com/v1), works with OpenAI / Anthropic compatible clients.BYOK,账单还在你手里——上游 Key 是你的,token 账单也在你原来的账户里,网关不转售 token,没有动力多发。
实测省 70.8%——170 次真实 Codex 请求,2318 万→677 万(客户端口径,非账单口径,测试方法)。
一个地址接所有模型——路由、故障自动切换、连接复用都在一个网关 Key 后面。
专治上下文膨胀——网关侧折叠、摘要、去水;132/170 次请求省 70%–90%,上下文越大省得越多。
免费公测——500 次/天,60 次/分钟,无需信用卡。
隐私设计——租户隔离、哈希存密钥、服务端不落盘 prompt 内容、Key 随时换随时废。
3 分钟接好——CC Switch 加自定义供应商即用,OpenAI / Anthropic 兼容客户端同理。
Why RLS / 为什么用 RLS
Local only, zero upload — 100% on your machine, listens on
127.0.0.1by default, no telemetry, no auto-update checks.−91.3% reading tokens on fixed sample tasks (reference magnitude, not a promise; how it was measured).
Built for AI agents first — native MCP for Claude Code / Codex; humans get the same index via CLI.
Five tools, one flow —
search→outline→context→trace→handover; file + symbol + line first, deep-read only what matters.Fast — warm query p95 ~20ms, exact search ~3.74ms; unknown symbol to definition + callers in 3 calls ≈ 80ms (trial walkthrough).
Honest results — exact literal match, never guesses; fuzzy queries return zero, unsolved edges are reported as gaps, not invented.
Single exe, 8 structured languages — Rust, Python, JS/TS, Go, Java, C/C++, others text-searchable. Free trial, Windows x64 first.
纯本地,零上传——100% 跑在你机器上,默认只听
127.0.0.1,无遥测、无自动更新检查。固定任务省 91.3% 阅读 token(样例量级参考,非承诺,测试方法)。
为 AI Agent 而生——Claude Code / Codex 原生 MCP,人用 CLI,同一下标同一语义。
五个工具一条链——
search→outline→context→trace→handover,先给文件+符号+行号,只深读必要的。快——暖查询 p95 约 20ms,精确搜索约 3.74ms;真实 trial 里 3 次调用约 80ms 从未知符号到定义+调用者(实测报告)。
不编答案——只认字面精确查询,打错就如实返回零结果;解不出的关系报 gap,不硬编边。
单个 exe,8 门结构化语言——Rust、Python、JS/TS、Go、Java、C/C++,其余文本可搜。免费试用,Windows x64 首发。
Contact / 联系
- Public questions → GitHub Issues (hirouter / rls)
- Business → [email protected] (email) · YangJie_3192 (WeChat, 请备注来意)
公开问题走 Issues,商务合作走邮箱/微信。