ClawRouter：专为 AI Agent 设计的 LLM 路由框架，55+ 模型、92% 省钱、<1ms 本地决策

ClawRouter: Agent-Native LLM Router — 55+ Models, 92% Cost Savings, <1ms Local Routing

Tech-Experiment #AI Agent#LLM路由#开源#成本优化#多模型#Web3#USDC#TypeScript#工具
更新于
🇨🇳 中文

仓库：BlockRunAI/ClawRouter · TypeScript · MIT · 6673 stars 获奖：USDC Hackathon — Agentic Commerce Winner npm：@blockrun/clawrouter · 官网：blockrun.ai

一、从一个问题出发

现有的 LLM 路由工具——OpenRouter、LiteLLM、Martian、Portkey——都是为人类开发者设计的：注册账号、生成 API key、绑信用卡、在仪表板里手选模型。

这套流程，AI Agent 一步都做不到。

ClawRouter 的出发点就是这一句话：

“Agents can’t sign up for accounts. Agents can’t enter credit cards. Agents can only sign transactions.”

二、它做了什么

ClawRouter 是一个本地运行的 LLM 代理路由器。你启动它之后，它在 localhost:8402 开一个 OpenAI 兼容的端点，任何工具只要把 API 地址指向这里，就能无感接入 55+ 个模型，路由逻辑全部在本地执行，不到 1ms 出结果。

和同类工具相比，它同时达成了五个条件——目前只有它一家：

OpenRouterLiteLLMMartianPortkeyClawRouter
开源✗✓✗部分✓
智能路由手动选手动选智能（闭源）观测智能（开源）
本地运行✗✓✗✗✓
加密原生✗✗✗✗✓
Agent 就绪✗✗✗✗✓

三、路由机制：15维评分，四级分层

每个请求进来，路由引擎在本地对其做 15 个维度的评分（不调用任何外部 API），给请求打一个复杂度标签，然后选出那个层级里最便宜的可用模型。

四个复杂度层级 × 三种路由策略：

层级ECO（最省钱）AUTO（默认）PREMIUM（最高质量）
SIMPLEfree/mistral-large-3-675b（免费）gemini-2.5-flashkimi-k2.7
MEDIUMgemini-3.1-flash-lite（$0.25/$1.50）kimi-k2.7（$0.95/$4.00）gpt-5.3-codex
COMPLEXgemini-3.1-flash-litegemini-3.1-pro（$2/$12）claude-opus-4.8
REASONINGdeepseek-reasoner（$0.20/$0.40）deepseek-reasonerclaude-sonnet-4.6

价格：千 token 输入/输出，单位美元。

效果：混合均价 $2.05/M tokens，对比 Claude Opus 4.8 的 $25/M，节省 92%。

路由策略用 /model 命令切换：

/model free       # 全部走免费模型，$0
/model auto       # 默认，平衡质量与成本（省74-100%）
/model eco        # 最省钱（省95-100%）
/model premium    # 最高质量，不省钱

也可以直接钉死某个模型：/model grok、/model br-sonnet、/model claude-opus。

四、Agent 原生设计：三个关键

4.1 免注册，钱包就是身份

首次运行时，ClawRouter 本地生成一个钱包。这个钱包地址就是你的身份，不需要邮箱、不需要密码、没有账号系统。Agent 程序可以持有这个钱包，像签交易一样”登录”。

npx @blockrun/clawrouter
# 首次运行打印钱包地址，也是你的认证凭证

4.2 USDC 微支付，x402 协议

付费模型的计费走 x402 协议——HTTP 请求本身携带 USDC 支付指令，在 Base 或 Solana 上结算。Agent 无需预先绑定信用卡，只要钱包里有 USDC，每次调用自动扣费，精确到 token 级别。

充值 $5 可以跑几千次请求。8 个免费模型完全不需要充值。

4.3 8 个永久免费模型，无需注册

免费层包含真正有竞争力的模型：

  • Mistral Large 3（675B 参数）
  • Qwen3-Next 80B（262K 上下文）
  • Nemotron Omni（支持视觉）
  • 以及另外 5 个 NVIDIA 托管模型

这个层级不需要任何余额、不需要注册，npx @blockrun/clawrouter 一条命令就能跑起来。

五、接入两分钟

选项 A：独立代理（continue.dev / Cursor / 任何 OpenAI 兼容客户端）

# 启动本地代理
npx @blockrun/clawrouter

代理起在 http://localhost:8402，用任何 OpenAI SDK 指向它：

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8402", api_key="x402")
response = client.chat.completions.create(
    model="blockrun/auto",   # 智能路由
    messages=[{"role": "user", "content": "你好"}]
)

continue.dev 配置（~/.continue/config.yaml）：

models:
  - name: ClawRouter Auto
    provider: openai
    model: blockrun/auto
    apiBase: http://localhost:8402/v1/   # 注意末尾斜杠
    apiKey: x402

Cursor：设置 → Models → OpenAI-compatible，base URL 填 http://localhost:8402，API key 填 x402，model 填 blockrun/auto。

选项 B：OpenClaw Agent 插件（一行命令）

curl -fsSL https://blockrun.ai/ClawRouter-update | bash
openclaw gateway restart

安装脚本处理全部配置：注册插件、同步模型列表、写认证文件、初始化钱包。完成后 blockrun/auto 成为默认模型。

六、不只是文本路由

ClawRouter 还把图像、视频、语音电话都路由了进来：

图像生成（/cr-imagegen）

/cr-imagegen a misty mountain at dawn
/cr-imagegen --model banana-pro --size 4096x4096 abstract geometric art
模型提供商价格最大分辨率
nano-bananaGoogle Gemini Flash$0.05/张1024×1024
banana-proGoogle Gemini Pro$0.10/张4096×4096
gpt-image-2OpenAI$0.06/张1536×1024
seedreamByteDance$0.045/张2848×1600
zai/cogview-4Zhipu$0.015/张1440×1440

视频生成（/videogen）

/videogen --model seedance-2-fast --duration=5 a cat waving

支持 Seedance 2（ByteDance）、Sora 2（Azure）、Grok Video，5秒视频约 $0.42-$1.49。

AI 电话（/cr-call）

/cr-call +14155552671 "Hi, calling to confirm tomorrow's 3pm meeting"

Bland.ai 支持，$0.54 固定费用，支持最长 30 分钟，支持转录和录音。

七、与其他路由器的本质区别

ClawRouter 之前，“智能路由”的意思是：你在仪表板里配规则，平台帮你调 API。路由逻辑在对方服务器上跑，每次路由决策要走一次网络。

ClawRouter 的路由是完全本地的：

Request → 本地 15维评分器 → Tier → 最优模型 → 直接转发
         （不到1ms，无外部依赖）

没有中间服务器知道你在用什么模型、问了什么问题。

对 Agent 系统来说这很重要：Agent 需要高频调用 LLM，路由延迟如果是 50-100ms，在一个多步 Agent 里会累积成秒级延迟。<1ms 的本地路由把这个成本消除了。

八、适合什么场景

直接适合：

  • 有多步 AI Agent，每步的任务复杂度差异大（写摘要 vs 代码推理）
  • 跑大量 LLM 请求，想在不降低效果的前提下控制成本
  • 想让 Agent 在没有人工干预的情况下自主管理 LLM 调用
  • 需要接入多个模型提供商，不想维护多套 SDK

需要额外考虑：

  • 15维本地评分是静态规则，不如端到端训练的路由模型那么精准（Martian 宣称的那种）
  • 免费模型质量参差，关键任务建议用 /model premium
  • x402 + USDC 是加密支付，对不接触 Web3 的团队有上手成本
  • 视频/电话这些能力是按量付费的，需要提前预估成本

数据来源：GitHub BlockRunAI/ClawRouter，2026-07-24 采集。

© 2026 Author: Mycelium Protocol

🇬🇧 English

Repository: BlockRunAI/ClawRouter · TypeScript · MIT · 6,673 stars Award: USDC Hackathon — Agentic Commerce Winner npm: @blockrun/clawrouter · Homepage: blockrun.ai

1. Starting from a Problem

Every existing LLM routing tool — OpenRouter, LiteLLM, Martian, Portkey — was designed for human developers: sign up for an account, generate an API key, enter a credit card, and pick a model from a dashboard.

AI agents can’t do any of that.

ClawRouter opens with exactly this premise:

“Agents can’t sign up for accounts. Agents can’t enter credit cards. Agents can only sign transactions.”

2. What It Does

ClawRouter is a locally-running LLM proxy router. Once started, it opens an OpenAI-compatible endpoint at localhost:8402. Any tool that points its API address here gets transparent access to 55+ models — all routing logic runs locally in under 1ms, with zero external API calls.

Compared to alternatives, it’s the only one that simultaneously checks all five boxes:

OpenRouterLiteLLMMartianPortkeyClawRouter
Open source✗✓✗Partial✓
Smart routingManualManualSmart (closed)ObservabilitySmart (open source)
Runs locally✗✓✗✗✓
Crypto native✗✗✗✗✓
Agent ready✗✗✗✗✓

3. How Routing Works: 15 Dimensions, 4 Tiers

When a request comes in, the routing engine evaluates it across 15 dimensions locally (no external API calls), assigns a complexity tier, then picks the cheapest capable model for that tier.

Four complexity tiers × three routing strategies:

TierECO (cheapest)AUTO (default)PREMIUM (best quality)
SIMPLEfree/mistral-large-3-675b (FREE)gemini-2.5-flashkimi-k2.7
MEDIUMgemini-3.1-flash-lite ($0.25/$1.50)kimi-k2.7 ($0.95/$4.00)gpt-5.3-codex
COMPLEXgemini-3.1-flash-litegemini-3.1-pro ($2/$12)claude-opus-4.8
REASONINGdeepseek-reasoner ($0.20/$0.40)deepseek-reasonerclaude-sonnet-4.6

Prices: per million input/output tokens in USD.

Net result: blended average $2.05/M tokens vs Claude Opus 4.8 at $25/M — 92% savings.

Switch strategies with /model:

/model free       # free models only, $0
/model auto       # default — balanced quality and cost (74–100% savings)
/model eco        # maximum savings (95–100%)
/model premium    # best quality, no savings

Pin a specific model: /model grok, /model br-sonnet, /model claude-opus.

4. Agent-Native Design: Three Core Properties

4.1 No signup — your wallet is your identity

On first run, ClawRouter generates a wallet locally. That wallet address is your identity — no email, no password, no account system. An agent program can hold this wallet and “authenticate” by signing transactions, exactly as it would with any blockchain operation.

npx @blockrun/clawrouter
# First run prints your wallet address — your auth credential

4.2 USDC micropayments via x402

Paid model billing runs through the x402 protocol — HTTP requests carry USDC payment instructions, settling on Base or Solana. No credit card needs to be pre-linked. Agents deduct automatically per request, to the token level.

$5 of USDC covers thousands of requests. The 8 free models require zero balance.

4.3 8 forever-free models — no signup required

The free tier includes genuinely capable models:

  • Mistral Large 3 (675B parameters)
  • Qwen3-Next 80B (262K context)
  • Nemotron Omni (vision-capable)
  • 5 additional NVIDIA-hosted models

No balance needed, no registration — npx @blockrun/clawrouter is the entire setup.

5. Two-Minute Setup

Option A: Standalone proxy (continue.dev / Cursor / any OpenAI-compatible client)

# Start the local proxy
npx @blockrun/clawrouter

Proxy runs at http://localhost:8402. Point any OpenAI SDK at it:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8402", api_key="x402")
response = client.chat.completions.create(
    model="blockrun/auto",
    messages=[{"role": "user", "content": "Hello"}]
)

continue.dev config (~/.continue/config.yaml):

models:
  - name: ClawRouter Auto
    provider: openai
    model: blockrun/auto
    apiBase: http://localhost:8402/v1/   # trailing slash required
    apiKey: x402

Cursor: Settings → Models → OpenAI-compatible, base URL: http://localhost:8402, API key: x402, model: blockrun/auto.

Option B: OpenClaw Agent plugin (single command)

curl -fsSL https://blockrun.ai/ClawRouter-update | bash
openclaw gateway restart

The script handles everything: plugin registration, model list sync, auth profile, wallet initialization. After completion, blockrun/auto becomes your default model.

6. Beyond Text: Images, Video, Voice

ClawRouter routes image generation, video, and voice calls through the same local proxy:

Image generation (/cr-imagegen)

/cr-imagegen a misty mountain at dawn
/cr-imagegen --model banana-pro --size 4096x4096 abstract art
ModelProviderPriceMax Size
nano-bananaGoogle Gemini Flash$0.05/image1024×1024
banana-proGoogle Gemini Pro$0.10/image4096×4096
gpt-image-2OpenAI$0.06/image1536×1024
seedreamByteDance$0.045/image2848×1600
zai/cogview-4Zhipu$0.015/image1440×1440

Video generation (/videogen)

/videogen --model seedance-2-fast --duration=5 a cat waving

Supports Seedance 2 (ByteDance), Sora 2 (Azure), Grok Video. 5-second clips run ~$0.42–$1.49.

AI phone calls (/cr-call)

/cr-call +14155552671 "Hi, calling to confirm tomorrow's 3pm meeting"

Powered by Bland.ai, $0.54 flat per call, up to 30 minutes, with transcript and recording retrieval.

7. The Core Distinction from Other Routers

Before ClawRouter, “smart routing” meant: you configure rules in a dashboard, the platform calls the API for you. Routing logic runs on someone else’s server; every routing decision crosses a network.

ClawRouter’s routing is entirely local:

Request → Local 15-dimension scorer → Tier → Best model → Direct forward
         (sub-1ms, zero external dependencies)

No intermediate server knows what model you’re using or what you asked.

For agent systems this matters: agents call LLMs at high frequency. If each routing decision adds 50–100ms, that accumulates to seconds of latency in a multi-step agent pipeline. Sub-1ms local routing eliminates this overhead entirely.

8. When to Use It

Good fit:

  • Multi-step AI agents where task complexity varies significantly per step (summarization vs. code reasoning)
  • High-volume LLM workloads where you want to control costs without sacrificing output quality
  • Agents that need to operate autonomously without human-managed API keys
  • Multi-provider model access without maintaining separate SDKs

Worth thinking about:

  • The 15-dimension local scorer is rule-based — not as precisely calibrated as an end-to-end trained routing model (like Martian claims)
  • Free model quality varies; use /model premium for mission-critical tasks
  • x402 + USDC is crypto payment — some friction for teams without Web3 exposure
  • Image/video/voice capabilities are pay-as-you-go — estimate volume before committing

Data source: GitHub BlockRunAI/ClawRouter, collected 2026-07-24.

© 2026 Author: Mycelium Protocol

相关文章

Tech-Experiment

Postiz：33K stars 的开源 Agent 社媒调度工具，让 AI 替你运营 30+ 平台
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

OpenSquilla：用多模型路由挑战顶级大模型，花 1/9 的钱达到同等效果

💬 评论与讨论

使用 GitHub 账号登录后发表评论