OmniRoute 杀疯了:一个 localhost 地址,把 Claude Code、Codex、Cursor 全接上,每月白嫖 16 亿 token!OmniRoute Goes Wild: One Localhost Address Hooks Up Claude Code, Codex, and Cursor — 1.6 Billion Free Tokens a Month!


AI free token/AI免费积分

「省 95% token,每月 16 亿免费额度,GitHub 上最火的趋势项目。」

6 月 28 日,GrowthSchool 创始人 Vaibhav Sisinty 在 X 上甩出这么一条帖子,配了一张密密麻麻的 token 预算表。评论区瞬间沸腾——2.2 万人围观,400 多人直接收藏。

他说的这个工具叫OmniRoute。

Vaibhav Sisinty 的爆款推文(分段1)Vaibhav Sisinty 的爆款推文(分段2)Vaibhav Sisinty 的爆款推文(分段3)

▲ Vaibhav Sisinty 原帖:「我刚发现一个工具,能把 AI token 成本降低 95%」,2.2 万次查看,283 赞

开发者的账单噩梦

先说一个很多人不愿面对的现实:2026 年,AI 编码工具已经从"尝鲜"变成了"离不开"。

Claude Code 每月 $100 起步,Cursor Pro 一年 $192,Codex 按 token 计费用多少扣多少。一个重度开发者同时开着三四个工具,月底看账单直接心跳加速。

更难受的是额度焦虑。Claude Pro 用户刚进入心流状态,右上角弹出"已达今日使用上限"。Cursor 用户把 tab 补全关了又开、开了又关,生怕浪费在不值得的地方。

还有一笔隐形账——token 浪费。你以为发给模型的 prompt 都是有用信息?看看 Claude Code 一个 30 分钟会话的实际构成:git diff、测试日志、构建输出、浏览器快照、反复传入的上下文……这些"噪音"占据了 60% 到 90% 的 token,对最终代码质量几乎没有贡献。

你在为垃圾信息买单。

OmniRoute 瞄准的就是这三个痛点。

一个 URL 统治所有工具

OmniRoute 的核心理念简单到令人怀疑:装一次,改一个地址,所有 AI 编码工具共享同一个入口。

具体来说,它在本地启动一个网关服务,地址是http://localhost:20128/v1。Claude Code、Cursor、Codex CLI、Cline、Copilot、Continue——凡是支持 OpenAI 或 Anthropic 兼容协议的工具,把 Base URL 指向这个地址就行。

背后,OmniRoute 连接了231 个 LLM 提供商,其中 50 多个有免费额度,11 个永久免费。

GitHub 项目主页

▲ OmniRoute GitHub 仓库,7.6k stars,1.3k forks,268 个 release,MIT 开源协议

这个项目在 GitHub 上已经迭代到 v3.8.39,4530 次提交,270 多位贡献者。光看标签就知道它的野心——mcp、gemini、openai、cursor、copilot、codex、cline、claude、a2a、deepseek、claude-code、token-saver……恨不得把整个 AI 编码生态全覆盖。

它用开发者 diegosouzapw 的话说:Never Stop Coding(永远别停下来写代码)。

压缩引擎:真正的杀手锏

省钱的秘密藏在两个压缩引擎里:RTK和Caveman。

RTK专门对付工具输出里的"废话"。git diff 里 500 行没变的上下文?砍掉。Jest 测试报告的进度条和 ANSI 颜色码?清除。构建日志里重复了 20 遍的 warning?只留一条。它用 49 条 JSON 过滤规则,精准识别并裁剪 Claude Code、Cursor 等工具最常产生的冗余输出。

官方给了一个真实案例:一个 30 分钟的 Claude Code 会话,原始消耗118,000 tokens,RTK 处理后只剩23,900 tokens——节省 79.7%。

Caveman则处理自然语言部分。它用一套多语言规则包,去掉填充词、犹豫表达和重复上下文,同时保护代码块、URL、JSON、文件路径这些结构化数据。

官方的一个压缩对比:

"The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle..."

压缩后变成:

"New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."

69 个 token 变成 19 个,节省 72%,语义完全等价。

当 RTK 和 Caveman 叠加使用(官方称为 Stacked 模式),节省公式是:

combined = 1 - (1 - RTK节省率) × (1 - Caveman节省率)

按官方平均数据,RTK 节省 80%,Caveman 节省 46%,叠加后约89.2%。极端情况下(工具密集型会话),节省率可以到 95%。

这就是"省 95% token"这个数字的来源。激进吗?确实。但它有数学支撑,也有 fidelity gate(保真关卡)机制——如果压缩后语义偏移超过阈值,会自动退回原文。

16 亿免费 token:怎么算出来的?

这个数字是 OmniRoute 爆火的另一个核心卖点,也是最容易引发争议的地方。

项目维护者专门写了一份FREE_TIERS.md文档,解释计数方法论。他们用了一个叫pool-deduped(池化去重)的原则:多个免费提供商共享同一个后端池时,只计数一次,避免虚报。

FREE_TIERS.md 文档

▲ 官方免费额度文档,详细拆分了各级别 token 预算与计算方式

具体拆分:

  • 稳态每月约 15.4 亿 token(Documented recurring grant)
  • 首月可达约 21.5 亿(含注册赠送积分)
  • 永久无上限类别(Pollinations、Qoder、LongCat 等,真实可用但未计入主数字)

最大贡献者是 Mistral(约 10 亿),其次是 Gemini Flash 家族、Cerebras、Cloudflare AI 等。

值得注意的是,2026 年 6 月 17 日的版本做了一次"诚实修正"——Gemini Flash 家族从虚高的 4.62 亿修正为约 6000 万(pool-dedup 后),Cloudflare 从标称数字修正为实际约 3000 万。这种主动挤水分的做法,在开源项目里相当少见。

当然,这 15.4 亿是文档化上限,真实可用量受 rate limit、并发数和提供商状态影响。把它理解成"天花板"比"保证"更准确。

智能路由:17 种策略 + 三层容灾

OmniRoute 的路由引擎支持 17 种策略,核心逻辑是Combo(模型链)。

你可以设一条 Combo:优先用付费订阅额度 → 额度用完切低价 API → 再切免费池。模型 填auto,OmniRoute 会用 9 个因子实时评分(健康度、剩余配额、延迟、历史成功率、价格、上下文适配等),自动选最优提供商。

当某个提供商挂了或者限流,三层独立容灾机制启动:

  1. Circuit Breaker(熔断器):整个提供商级别隔离
  2. Connection Cooldown(连接冷却):单个 API key 限流时,其他 key 继续工作
  3. Model Lockout(模型锁定):只锁出问题的模型,同一提供商的其他模型不受影响

故障切换据称在10 毫秒以内完成。对开发者来说,正在跑的 agent 任务不会中断,甚至感知不到背后换了一个提供商。

5 分钟上手

Vaibhav 在跟帖里贴了完整的设置指南,简化到三步:

设置指南推文(分段1)设置指南推文(分段2)设置指南推文(分段3)

▲ 5 分钟设置 OmniRoute:安装 → 连接免费提供商 → 指向一个 URL

第一步:安装

npm install -g omniroute omniroute 

Dashboard 自动打开在localhost:20128。

第二步:连接免费提供商

Dashboard → Providers → 点击 Kiro AI 或 OpenCode Free。不需要信用卡,不需要注册。Pollinations 甚至连 API key 都不用。

第三步:把工具指向一个 URL

所有编码工具的 Base URL 改成http://localhost:20128/v1,Model 填auto,API Key 从 Dashboard 复制。

Claude Code 用户改~/.claude/settings.json:

{   "env": {     "ANTHROPIC_BASE_URL": "http://localhost:20128"   } } 

Codex CLI 用户设环境变量:

export OPENAI_BASE_URL="http://localhost:20128" 

v3.8 之后还加了一键命令:omniroute setup claude、omniroute setup cursor,连手动改配置都省了。

冷静一下:风险与争议

数字好看归好看,但上手之前有几件事必须知道。

ToS 灰色地带。部分免费提供商(Kiro、Gemini CLI、Antigravity)的服务条款明确禁止通过第三方代理访问。OmniRoute 的 FREE_TIERS.md 里专门列了一张 ToS 标记表,但用不用、风险自负。

压缩质量的边界。89% 的平均节省率来自工具密集型会话。如果你的 prompt 本身就很精简(比如直接贴一段代码问"哪里有 bug"),压缩带来的收益可能只有 15%-20%。而在极端复杂的上下文里,ultra 模式有概率引入语义偏移——虽然有 fidelity gate,但官方也承认需要用户自测 edge case。

"黑盒切换"的调试难题。auto 模式在后台自动换提供商和模型,表面上一切正常,但有 Reddit 用户反映遇到过"模型切换后 system prompt 行为微漂移"的情况。当你的 agent 任务出了诡异结果,定位问题会变得更复杂。

社区争议。Reddit 上有帖子对比 OmniRoute 与早期项目 9router 的代码和 UI,相似度很高。OmniRoute 被认为是"受启发后重构 + 增加了 MCP/A2A/更强容灾",但关于"fork 还是重写"的争论并未平息。

RoundtableSpace 的总结帖(分段1)RoundtableSpace 的总结帖(分段2)RoundtableSpace 的总结帖(分段3)

▲ @RoundtableSpace 总结帖:1.7 万次查看,148 赞——「所有免费层叠加,每个 token 发挥最大价值」

为什么偏偏在这个节点火了?

时机很关键。

2026 年上半年,AI 编码工具从"可选"变成了"标配"。Claude Code、Cursor、Codex 三足鼎立,开发者从"试用一个"变成了"同时依赖三四个"。多工具订阅的账单压力集中爆发。

与此同时,免费提供商的生态也在悄悄成熟——Pollinations 提供无需 key 的 GPT-5/Claude 访问,Qoder 给出无限量的 Kimi-K2 和 DeepSeek-R1,LongCat 每天 5000 万 token 的 Flash-Lite。这些散落各处的免费资源,需要有人把它们整合起来。

OmniRoute 做的就是这件事。再配上"95% 省 token + 16 亿免费 + 3 步设置"这组极具冲击力的数字,加上 AI 内容账号的模板化推广——一切条件都齐了。

写在最后

OmniRoute 代表的趋势比项目本身更值得关注:Token 优化正在成为 AI 编码的第二战场。

过去两年,开发者在"选哪个模型"上花了无数精力。现在,当模型能力趋于同质化,竞争焦点转向了基础设施层——谁能用更少的 token 完成同样的任务,谁能在多个提供商之间无缝切换,谁能把分散的免费额度拧成一股绳。

OmniRoute 用"本地网关 + 压缩引擎 + 智能路由"这套组合拳给出了一个答案。7600 多颗星、270 多位贡献者、268 个 release——这个项目的迭代速度说明社区真的有需求。

但魔法终归有代价。免费池的 ToS 风险、压缩的保真边界、自动路由的可观测性——每一条都需要你亲自验证。

建议的姿势是:先装上跑两天,测自己 workflow 里的压缩质量和 fallback 行为。生产环境的关键路径,留一条付费 API 兜底。

毕竟,免费的午餐吃得开心,但别把锅都扔了。

 

"Save 95% on tokens, 1.6 billion free monthly quota, the hottest trending project on GitHub."

On June 28, GrowthSchool founder Vaibhav Sisinty posted this on X, along with a dense token budget table. The comments section instantly exploded—22,000 people viewed it, and over 400 directly bookmarked it.

The tool he was talking about is called OmniRoute.

▲ Vaibhav Sisinty's original post: "I just found a tool that cuts AI token costs by 95%", 22,000 views, 283 likes

The developer's billing nightmare

Let's start with a reality many people don't want to face: by 2026, AI coding tools have gone from "trying them out" to "can't live without them."

Claude Code starts at $100/month, Cursor Pro costs $192/year, and Codex charges per token—the more you use, the more you pay. A heavy developer running three or four tools at once will feel their heart rate spike when the monthly bill arrives.

What's even worse is quota anxiety. Claude Pro users finally enter a flow state, and a pop-up appears in the top-right corner: "Daily usage limit reached." Cursor users toggle tab completion off and on, off and on, terrified of wasting it on something unworthy.

Then there's the invisible cost—token waste. Think everything you send to the model as a prompt is useful? Look at the actual composition of a 30-minute Claude Code session: git diffs, test logs, build output, browser snapshots, repeatedly passed-in context... This "noise" takes up 60% to 90% of tokens while contributing almost nothing to the final code quality.

You're paying for junk information.

OmniRoute targets exactly these three pain points.

One URL to rule all tools

OmniRoute's core concept is so simple it makes you skeptical: install once, change one address, and all AI coding tools share the same entry point.

Specifically, it starts a local gateway service at http://localhost:20128/v1. Claude Code, Cursor, Codex CLI, Cline, Copilot, Continue—any tool that supports OpenAI or Anthropic-compatible protocols just points its Base URL to this address.

Behind the scenes, OmniRoute connects to 231 LLM providers, over 50 of which offer free quotas, and 11 are permanently free.

▲ OmniRoute GitHub repository, 7.6k stars, 1.3k forks, 268 releases, MIT open-source license

This project has iterated to v3.8.39 on GitHub, with 4,530 commits and 270+ contributors. Just looking at the tags reveals its ambition—mcp, gemini, openai, cursor, copilot, codex, cline, claude, a2a, deepseek, claude-code, token-saver... it's trying to cover the entire AI coding ecosystem.

In the words of developer diegosouzapw: Never Stop Coding.

The compression engine: the real killer feature

The secret to saving money lies in two compression engines: RTK and Caveman.

RTK targets the "fluff" in tool output. A git diff with 500 lines of unchanged context? Cut it. Jest test report progress bars and ANSI color codes? Stripped out. The same warning repeated 20 times in a build log? Keep only one. It uses 49 JSON filtering rules to precisely identify and trim the redundant output most commonly generated by tools like Claude Code and Cursor.

The official docs give a real case: a 30-minute Claude Code session originally consumed 118,000 tokens; after RTK processing, only 23,900 tokens remained—a 79.7% saving.

Caveman handles the natural language side. It uses a multilingual rule package to strip filler words, hesitant expressions, and repeated context, while protecting structured data like code blocks, URLs, JSON, and file paths.

An official compression comparison:

"The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle..."

After compression, it becomes:

"New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."

69 tokens became 19, a 72% saving, with semantically equivalent meaning.

When RTK and Caveman are stacked (officially called Stacked mode), the savings formula is:

combined = 1 - (1 - RTK savings rate) × (1 - Caveman savings rate)

Based on official average data, RTK saves 80% and Caveman saves 46%; stacked together, it's about 89.2%. In extreme cases (tool-intensive sessions), the savings rate can reach 95%.

That's where the "save 95% on tokens" figure comes from. Aggressive? Absolutely. But it has mathematical backing, and there's also a fidelity gate mechanism—if the compressed output drifts beyond a threshold semantically, it automatically falls back to the original text.

1.6 billion free tokens: how is that calculated?

This number is another core selling point behind OmniRoute's explosion in popularity, and also the most likely source of controversy.

The project maintainers wrote a dedicated FREE_TIERS.md document explaining the counting methodology. They use a principle called pool-deduped: when multiple free providers share the same backend pool, it's counted only once to avoid inflated numbers.

▲ Official free quota documentation, with a detailed breakdown of token budgets and calculation methods at each tier

Specific breakdown:

Steady-state monthly quota of approximately 1.54 billion tokens (Documented recurring grant)

First month can reach approximately 2.15 billion (including registration bonus credits)

Permanent unlimited categories (Pollinations, Qoder, LongCat, etc.—actually usable but not included in the main figure)

The largest contributor is Mistral (about 1 billion), followed by the Gemini Flash family, Cerebras, Cloudflare AI, and others.

Notably, the June 17, 2026 version made an "honesty correction"—the Gemini Flash family was revised from an inflated 462 million to about 60 million (after pool-dedup), and Cloudflare was revised from its nominal figure to an actual ~30 million. This kind of proactively squeezing out the inflation is quite rare in open-source projects.

Of course, this 1.54 billion is the documented ceiling; the actually usable amount is affected by rate limits, concurrency, and provider status. It's more accurate to understand it as a "ceiling" than a "guarantee."

Smart routing: 17 strategies + three-layer failover

OmniRoute's routing engine supports 17 strategies, with the core logic being Combo (model chains).

You can set up a Combo: prioritize paid subscription quota → switch to low-cost APIs when quota runs out → then switch to the free pool. Set the model to auto, and OmniRoute will score in real time using 9 factors (health, remaining quota, latency, historical success rate, price, context fit, etc.) to automatically pick the best provider.

When a provider goes down or gets rate-limited, a three-layer independent failover mechanism kicks in:

Circuit Breaker: isolates at the entire provider level

Connection Cooldown: when a single API key is rate-limited, other keys keep working

Model Lockout: only locks the problematic model; other models from the same provider remain unaffected

Failover is reportedly completed within 10 milliseconds. For developers, running agent tasks won't be interrupted—they may not even notice a provider switch happened in the background.

Get started in 5 minutes

Vaibhav posted a complete setup guide in the replies, simplified down to three steps.

▲ Set up OmniRoute in 5 minutes: install → connect free providers → point to one URL

Step 1: Install

The Dashboard automatically opens at localhost:20128.

Step 2: Connect free providers

Dashboard → Providers → click Kiro AI or OpenCode Free. No credit card needed, no registration needed. Pollinations doesn't even require an API key.

Step 3: Point your tools to one URL

Change the Base URL of all coding tools to http://localhost:20128/v1, set Model to auto, and copy the API Key from the Dashboard.

Claude Code users edit ~/.claude/settings.json:

Codex CLI users set environment variables:

After v3.8, one-command setups were added: omniroute setup claude, omniroute setup cursor—even manual config editing is no longer needed.

Let's stay calm: risks and controversies

The numbers look great, but there are a few things you need to know before getting started.

ToS gray area. Some free providers (Kiro, Gemini CLI, Antigravity) explicitly prohibit third-party proxy access in their terms of service. OmniRoute's FREE_TIERS.md includes a ToS flag table, but whether you use them is your own risk.

The limits of compression quality. The 89% average savings rate comes from tool-intensive sessions. If your prompts are already lean (for example, directly pasting a piece of code and asking "where's the bug"), the compression benefit may only be 15%-20%. And in extremely complex contexts, ultra mode has a chance of introducing semantic drift—although the fidelity gate exists, the official docs also admit that users need to test edge cases themselves.

The debugging challenge of "black-box switching." Auto mode silently switches providers and models in the background, and everything looks normal on the surface, but some Reddit users have reported cases where "the system prompt behavior drifts slightly after a model switch." When your agent task produces weird results, pinpointing the problem becomes more complicated.

Community controversy. There are Reddit posts comparing OmniRoute's code and UI with the earlier project 9router, and the similarity is high. OmniRoute is considered "inspired, then rewritten + added MCP/A2A/stronger failover," but the debate over "fork vs. rewrite" hasn't died down.

▲ @RoundtableSpace's summary post: 17,000 views, 148 likes—"Stack all the free tiers, get maximum value out of every token"

Why did it blow up precisely at this point?

Timing is crucial.

In the first half of 2026, AI coding tools went from "optional" to "standard." Claude Code, Cursor, and Codex formed a three-way standoff, and developers went from "trying one" to "depending on three or four at once." The bill pressure from multiple tool subscriptions hit all at once.

At the same time, the ecosystem of free providers was quietly maturing—Pollinations offers keyless GPT-5/Claude access, Qoder offers unlimited Kimi-K2 and DeepSeek-R1, and LongCat offers 50 million tokens per day of Flash-Lite. These scattered free resources needed someone to consolidate them.

That's exactly what OmniRoute does. Pair that with the highly impactful set of numbers—"95% token savings + 1.6 billion free + 3-step setup"—plus template-style promotion from AI content accounts, and all the conditions were in place.

Final thoughts

The trend OmniRoute represents is more worth paying attention to than the project itself: token optimization is becoming the second battlefield in AI coding.

Over the past two years, developers have spent endless effort on "which model to choose." Now that model capabilities are becoming homogeneous, the competitive focus has shifted to the infrastructure layer—who can complete the same task with fewer tokens, who can switch seamlessly between multiple providers, and who can weave scattered free quotas into a single rope.

OmniRoute offers one answer with the combination of "local gateway + compression engine + smart routing." More than 7,600 stars, 270+ contributors, and 268 releases—this project's iteration speed shows the community genuinely has a need.

But magic always comes at a cost. The ToS risks of the free pool, the fidelity limits of compression, and the observability of automatic routing—every one of these needs to be verified by you personally.

The recommended approach: install it first and run it for two days, testing the compression quality and fallback behavior in your own workflow. For critical paths in production, keep a paid API as a safety net.

After all, a free lunch is nice, but don't throw away the pot.

Vaibhav Sisinty 的爆款推文(分段1)

Vaibhav Sisinty 的爆款推文(分段2)

Vaibhav Sisinty 的爆款推文(分段3)

GitHub 项目主页

FREE_TIERS.md 文档

设置指南推文(分段1)

设置指南推文(分段2)

设置指南推文(分段3)

RoundtableSpace 的总结帖(分段1)

RoundtableSpace 的总结帖(分段2)

RoundtableSpace 的总结帖(分段3)

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry