OmniRoute：290 个 AI 提供商、一个端点、永不限流——免费 AI 网关的暴力美学

OmniRoute: 290 AI Providers, One Endpoint, Never Hit Limits — The Free AI Gateway

Tech-Experiment #AI网关#开源#Claude Code#免费AI#路由#MCP#token压缩#开发工具
更新于
🇨🇳 中文

仓库：diegosouzapw/OmniRoute · TypeScript · MIT · 27824 stars 最新版本：v3.8.48（2026-07-13） 创建时间：2026-02-13（5个月，500+ 贡献者） 官网：omniroute.online · 端点：http://localhost:20128/v1

一、它解决什么问题

凌晨 2 点在写代码，Claude Code 的配额用完了。

你的选择通常是：等重置、换个 API key、去 OpenRouter 注册一个账号。这个流程每次都要打断你的思路。

OmniRoute 的答案是：你根本不应该遇到这个问题。

它在本地运行一个 AI 网关，把你能用的所有 AI 来源——订阅账号、API key、免费 tier、零配置免费提供商——组合成一个端点，按你配置的策略自动故障转移。配额用完了，自动切到下一个。你的工具感知不到任何变化。

二、规模

  • 290+ 提供商，500+ 模型（Kimi、Claude、GPT、Gemini、GLM、DeepSeek、MiniMax……）
  • 90+ 免费提供商，其中 40+ 是永久免费（不需要注册，不需要 key）
  • ~1.53B 免费 tokens/月：聚合 43 个提供商池 / 460+ 模型的文档化免费 tier，在 dashboard 实时显示剩余量
  • 零配置就能用：安装后 auto 模式立即可用，内置接了 OpenCode Free 和 Felo 等无 key 提供商
# 安装后无需任何配置——auto 立即响应
curl http://localhost:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Hello!"}]}'

支持的 AI coding 工具：Claude Code、Codex、Cursor、OpenCode、Cline、Copilot，以及所有 OpenAI API 兼容的客户端。

三、4 层自动 Fallback

OmniRoute 把所有来源按优先级分成 4 层：

Tier 1 订阅账号 ─── Claude Code / Codex / Copilot（最贵但最快）
         ↓ 配额用完
Tier 2 API Key ────── DeepSeek / Groq / xAI（付费 API，便宜）
         ↓ 预算触碰
Tier 3 廉价 ──────── GLM($0.5) / MiniMax($0.2)
         ↓ 预算触碰
Tier 4 免费 ──────── Kiro / Qoder / Pollinations（零成本）

在你的工具看来，这始终是同一个端点 localhost:20128/v1，模型名 auto。切换对工具完全透明。

四、19 种路由策略

这是 OmniRoute 和简单代理最大的工程差距。

零配置的 auto 变体：

模型 ID优化目标
auto平衡（LKGP——粘性上次成功路径）
auto/coding代码生成质量优先
auto/fast最低延迟优先
auto/cheap最低 token 成本优先
auto/offline最多剩余配额优先
auto/smart质量优先 + 10% 探索发现更好模型

可自定义 Combo 的 19 种策略（选几个有意思的）：

  • context-relay：跨目标传递对话上下文——长对话自动切 provider 时保持连贯
  • cache-optimized：把同一个可复用 prompt 前缀锁定在同一个账号，最大化 prompt cache 命中率
  • lkgp：Last-Known-Good Path，粘性到上次成功的目标
  • fusion：把请求扇出到多个模型，用一个 judge 把多个答案合成一个——用于高质量任务
  • pipeline：链式执行，上一步输出是下一步输入——用于多步骤工作流
  • quota-share：一个订阅账号在团队成员间按权重公平分配配额（weight 50/30/20），支持 hard/soft/burst 模式

五、RTK + Caveman 压缩

OmniRoute 在发请求之前对 token 做压缩，自称节省 15-95%，工具密集型 session 平均 ~89%。

这个数字如果属实，意义非常大——特别是对 Claude Code 这样工具调用很多的场景：每次工具调用的参数和返回结果可能占大量 token，如果能压缩掉 89%，等效于你的配额变成 9 倍。

两层叠加：RTK（基于工具调用特征的压缩）和 Caveman（通用文本压缩）。

六、MCP Server + A2A

OmniRoute 不只是一个代理，它自己也是一个 MCP 服务端：

  • 104 个 MCP 工具，3 种传输方式，31 个 scope
  • A2A（Agent-to-Agent）：6 个 skill，JSON-RPC 2.0——Agent 可以互相调用

这意味着：Claude Code 或 Cursor 可以把 OmniRoute 当成一个 MCP server 接入，直接获得 OmniRoute 的路由能力和工具库。

七、3 层弹性

三个独立的自愈层，对应三个粒度的故障：

Layer 1 — Provider 断路器（整个提供商级别） 只在 408/5xx 时触发，阈值是 OAuth 3次/API key 5次/本地 2次，重置 60s/30s/15s 后进入 HALF-OPEN 探测，恢复后继续路由。

Layer 2 — Connection 冷却（单 key/账号级别） 429 时遵守 Retry-After，基础冷却 5s(OAuth)/3s(API key)，指数退避 ×2，同池其他 key 继续服务。

Layer 3 — Model 锁定（单模型级别） 一个模型的 429 或 404 只锁定这个模型，不影响同 provider 的其他模型。

八、数字

  • 27,824 stars，3,655 forks
  • 2026-02-13 创建——5个月做到 2.7 万 stars
  • 500+ 贡献者，文档已翻译成 43 种语言
  • 175 open issues，61 open PRs
  • v3.8.48——版本号说明已经迭代了几百次
  • npm 包 omniroute，Docker Hub 镜像，Electron 桌面应用，PWA

九、判断

OmniRoute 踩在了一个真实的开发者痛点上：配额限流打断工作流。这个问题在 AI 工具高度普及的 2026 年越来越普遍——Claude Code、Codex、Cursor 同时开着，配额消耗比以前快 10 倍。

5 个月 2.7 万 stars，500+ 贡献者，这个增速不是靠一次 HN 首页驱动的，是持续的真实需求驱动。

几个值得关注的细节：

  • Kimi（Moonshot AI）是 Founding Open Source Friend：Kimi K3 的 API credits 支持了 OmniRoute 的 merge validation pipeline——每个 PR 合并前都用 Kimi K3 做 AI review。这是一个务实的开源赞助关系：不是 logo 赞助，是真实集成进 CI/CD。

  • RTK+Caveman 的 89% 压缩率：README 里有 CI gate 保证文档数字和代码实际输出一致，不是随意填写的。如果在工具密集型 session 里真的能平均省 89%，这是比”多几个免费提供商”更重要的价值。

  • Quota-Share 机制：对于用同一个 Claude Max 订阅开着 4 个 Claude Code 窗口的开发者来说，Quota-Share 可以让每个窗口有公平的配额份额而不是互相抢。

和 Osaurus（本地哈尼斯）、OpenWorker（桌面工作助手）相比，OmniRoute 的定位是更底层的基础设施：它不关心你用 AI 做什么，只保证你的 AI 调用永远有地方可去。

不确定的地方：

  • 290 个提供商里有多少是真正可用的（不是只在 provider list 里但实际要翻墙或有地区限制）？
  • RTK+Caveman 的压缩对不同类型的任务效果差异有多大？非工具密集型场景是否值得开？
  • 免费 tier 的 1.53B tokens 背后的 QPS 限制怎么分布？高峰时段是否会遇到整体降速？

总结：如果你在用 Claude Code / Cursor / Codex，OmniRoute 是值得 5 分钟试一次的基础设施工具——最坏情况是白装了，最好情况是你的配额有效扩大了 10 倍。

数据来源：GitHub diegosouzapw/OmniRoute，2026-07-24 采集。

© 2026 Author: Mycelium Protocol

🇬🇧 English

Repository: diegosouzapw/OmniRoute · TypeScript · MIT · 27,824 stars Latest Release: v3.8.48 (2026-07-13) Created: 2026-02-13 (5 months, 500+ contributors) Homepage: omniroute.online · Endpoint: http://localhost:20128/v1

1. What Problem It Solves

It’s 2 AM. You’re deep in a coding session and Claude Code’s quota runs out.

Your usual options: wait for the reset, swap in a different API key, sign up for an OpenRouter account. Every path breaks your flow.

OmniRoute’s answer: you shouldn’t hit this problem in the first place.

It runs a local AI gateway that combines every AI source you have access to — subscriptions, API keys, free tiers, zero-config keyless providers — into one endpoint, automatically failing over according to whatever strategy you configure. When one quota runs dry, it silently routes to the next. Your tools see nothing change.

2. The Scale

  • 290+ providers, 500+ models (Kimi, Claude, GPT, Gemini, GLM, DeepSeek, MiniMax…)
  • 90+ free providers, 40+ free forever (no signup, no key required)
  • ~1.53B free tokens/month: aggregates documented free tiers from 43 provider pools / 460+ models, shown live on the dashboard
  • Zero-config from install: auto mode responds immediately — OpenCode Free and Felo are wired in out of the box with no keys
# After install, no credentials needed — auto responds immediately
curl http://localhost:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Hello!"}]}'

Supported AI coding tools: Claude Code, Codex, Cursor, OpenCode, Cline, Copilot, and any OpenAI-compatible client.

3. The 4-Tier Auto-Fallback Cascade

OmniRoute ranks all sources into four priority tiers:

Tier 1 Subscription ── Claude Code / Codex / Copilot (best quality)
        ↓ quota exhausted
Tier 2 API Key ──────── DeepSeek / Groq / xAI (paid, inexpensive)
        ↓ budget hit
Tier 3 Cheap ──────── GLM($0.5) / MiniMax($0.2)
        ↓ budget hit
Tier 4 Free ────────── Kiro / Qoder / Pollinations (zero cost)

From your tool’s perspective, this is always one endpoint at localhost:20128/v1, model name auto. Provider switches are completely transparent.

4. 19 Routing Strategies

This is the biggest engineering gap between OmniRoute and simple proxies.

Zero-config auto variants:

Model IDOptimizes for
autoBalanced (LKGP — sticky to last successful path)
auto/codingCode generation quality first
auto/fastLowest latency first
auto/cheapLowest per-token cost first
auto/offlineMost remaining quota first
auto/smartQuality-first + 10% exploration to discover better models

Notable strategies for custom Combos:

  • context-relay: Hands off conversation context across targets — long conversations stay coherent when switching providers mid-session
  • cache-optimized: Pins the same reusable prompt prefix to the same account to maximize prompt-cache hit rate
  • lkgp: Last-Known-Good Path — sticky to the last successful target
  • fusion: Fans out to a panel of models, a judge synthesizes one answer — for high-quality critical tasks
  • pipeline: Chain execution where each target’s output feeds the next — for multi-step workflows
  • quota-share: Distributes one subscription’s quota fairly across team members by weight (e.g. 50/30/20), with hard/soft/burst modes

5. RTK + Caveman Compression

OmniRoute compresses tokens before sending requests. The claimed savings: 15–95%, averaging ~89% on tool-heavy sessions.

If this holds in practice, the implications are significant — especially for Claude Code and Cursor, where tool calls dominate token count: every tool call’s parameters and return values can consume large amounts of context. If 89% of that compresses away, your effective quota becomes roughly 9× larger.

Two layers stacked: RTK (compression tuned for tool-call patterns) and Caveman (general-purpose text compression). The README notes a CI gate (check:docs-counts) that fails the build if the documented numbers drift from what the code actually produces — these aren’t figures someone typed by hand.

6. MCP Server + A2A

OmniRoute isn’t just a proxy — it also runs as an MCP server itself:

  • 104 MCP tools, 3 transport types, 31 scopes
  • A2A (Agent-to-Agent): 6 skills, JSON-RPC 2.0 — agents can call each other

This means Claude Code or Cursor can connect to OmniRoute as an MCP server and gain access to OmniRoute’s routing capabilities and tool library directly.

7. Three-Layer Resilience

Three independent self-healing layers, each targeting a different failure granularity:

Layer 1 — Provider circuit breaker (entire provider level) Trips only on 408/5xx: thresholds at OAuth 3× / API key 5× / local 2×. Resets into HALF-OPEN probe at 60s/30s/15s, then lazy recovery. While OPEN, the combo reroutes to the next provider.

Layer 2 — Connection cooldown (single key/account level) 429 responses honor Retry-After. Base cooldown 5s (OAuth) / 3s (API key), exponential ×2 backoff. A cooling key is skipped; sibling keys keep serving.

Layer 3 — Model lockout (single model level) A 429 or 404 on one model locks only that model — never the whole connection or provider.

8. Numbers

  • 27,824 stars, 3,655 forks
  • Created 2026-02-13 — 2.7K stars in 5 months
  • 500+ contributors, docs translated into 43 languages
  • 175 open issues, 61 open PRs
  • v3.8.48 — the version number signals hundreds of iterations already
  • Available as npm package, Docker Hub image, Electron desktop app, and PWA

9. Assessment

OmniRoute targets a real developer pain point: rate-limit interruptions breaking flow. This is increasingly common in 2026 as AI tools proliferate — running Claude Code, Codex, and Cursor simultaneously burns quota roughly 10× faster than before.

2.7K stars in 5 months with 500+ contributors isn’t driven by one HN front page. That’s sustained real-world demand.

A few details worth noting:

  • Kimi (Moonshot AI) is the Founding Open Source Friend: Kimi K3 API credits power OmniRoute’s merge validation pipeline — every PR gets AI-reviewed by Kimi K3 before shipping. This is a practical sponsorship relationship, not logo placement: it’s actually integrated into CI/CD.

  • The 89% RTK+Caveman compression figure: The README includes a CI gate that fails the build if the documented numbers drift from what the code computes. These aren’t marketing estimates. If they hold in tool-heavy sessions, this is more valuable than access to more free providers.

  • Quota-Share: For developers running 4 Claude Code windows against one Claude Max subscription, Quota-Share can give each window a fair share of the quota rather than having them compete.

Compared with Osaurus (local Mac harness) and OpenWorker (desktop work assistant), OmniRoute sits at a lower layer of the stack: it doesn’t care what you do with AI, only that your AI calls always have somewhere to go.

Open questions:

  • Of the 290 providers, how many are actually accessible without VPN or regional restrictions?
  • How does RTK+Caveman compression vary by task type? Is it worth enabling for non-tool-heavy sessions?
  • How does QPS distribute across the 1.53B free tokens? Will peak-hour usage hit aggregate slowdowns?

Bottom line: If you use Claude Code, Cursor, or Codex, OmniRoute is worth a 5-minute trial as infrastructure. Worst case: you installed something you don’t need. Best case: your effective quota just expanded by 10×.

Data source: GitHub diegosouzapw/OmniRoute, collected 2026-07-24.

© 2026 Author: Mycelium Protocol

相关文章

Tech-Experiment

ego lite：为 Claude Code 设计的并行浏览器，一次 JS 调用完成整个任务
Tech-Experiment

FireRed-OpenStoryline：用对话剪视频，把剪辑工作流存成可复用的 Skill
Tech-Experiment

BrowserAct：给 AI Agent 装上真实浏览器，4500 stars，10M+ CAPTCHA 已解决

💬 评论与讨论

使用 GitHub 账号登录后发表评论