ClawBuddy 🦞

Cost-optimization buddy for OpenClaw. Make your AI assistant cheaper over time.

ClawBuddy is a cost-optimization plugin for OpenClaw. It works as a transparent proxy that intercepts OpenClaw API calls, collects data in the background, validates cheap model capabilities, and auto-generates Skills — gradually migrating expensive large model calls to cheaper small models without sacrificing task quality.

How It Works

User → OpenClaw → ClawBuddy proxy (:18790) → Primary API (e.g. glm-5-turbo)
                         ↓
                ┌──────────────────────┐
                │   Background (async)  │
                │                      │
                │  Mode 1: Data collection   │
                │  Mode 2: Mock validation   │ ← cheap model (e.g. qwen3.5:9b) + skills
                └──────────────────────┘

ClawBuddy launches a local OpenAI-compatible proxy server and modifies the OpenClaw config to point API requests to it. Fully transparent to OpenClaw — it just sees a normal model provider.

Two Modes

📦 Mode 1: Data Distiller

Pure data collection, zero intrusion.

All requests forwarded to the primary model as usual, zero UX change
Automatically collects complete input/output for every request
Chains multi-turn requests from the same session into complete task traces
Auto-labels each task with a cheap model (e.g. "file organization", "code refactoring")
Export as JSONL / LLaMA-Factory format for fine-tuning your own small model

Best for: Users with GPU resources who want to fundamentally improve small model capabilities through training.

🧪 Mode 2: Mock Validator + Skill Generator (Recommended)

Sandbox validation + auto Skill generation, training-free cost reduction.

Core loop:

Use primary model (glm-5-turbo) for daily tasks
    ↓ Background collects task traces
Run the same task with cheap model (qwen3.5:9b) in a mock sandbox
    ↓ Primary model judges the result
pass → Accumulate pass records
fail → Analyze failure → Auto-generate OpenClaw Skill → Re-validate with Skill
    ↓ Task type accumulates N passes
Dashboard notification → "Poetry writing" validated, suggest switching to cheap model
    ↓ User enables offload with one click

The primary model is only called at three low-frequency points: generating mock environments, judging results, and generating Skills. Daily labeling is handled entirely by the cheap model.

Best for: Users without GPUs or who prefer prompt engineering over model training.

Quick Start

Installation

git clone https://github.com/haochen2115/clawbuddy.git
cd clawbuddy
npm install
npm run build
npm link  # Makes 'clawbuddy' command available globally

Usage

One command to start, one to stop:

# Start (auto-backup OpenClaw config, inject proxy, launch services)
clawbuddy start

# Stop (Ctrl+C or run from another terminal)
clawbuddy stop

That's it. After starting, use OpenClaw as usual. Open http://localhost:18791 for the Dashboard.

start automatically:

Backs up OpenClaw config (first launch only)
Injects proxy provider into OpenClaw (OpenClaw hot-reloads, no restart needed)
Launches proxy server + dashboard

stop (or Ctrl+C) automatically:

Stops all services
Precisely restores OpenClaw config (only removes proxy-related fields, preserves other changes)

Configuration

Default config works out of the box, but you can customize via CLI flags:

# Specify mode (default: mode2)
clawbuddy start --mode mode1

# Specify cheap model
clawbuddy start --cheap-provider ollama --cheap-model qwen3.5:9b

# Or edit config file
clawbuddy config set mode mode1
clawbuddy config set cheap_api.model qwen3.5:9b
clawbuddy config show

Export Training Data

# Export all
clawbuddy export --format jsonl --output ./training_data.jsonl

# Filter by task label
clawbuddy export --label "file organization" --output ./file_org.jsonl

# Export as LLaMA-Factory format
clawbuddy export --format llamafactory --output ./dataset/

Manage Offload (Mode 2)

# Check status
clawbuddy status

# Enable / disable offload
clawbuddy offload enable "poetry writing"
clawbuddy offload disable "poetry writing"
clawbuddy offload list

Dashboard

Local Web Dashboard at http://localhost:18791 with 6 panels:

Panel	Function
Overview	Mode, task counts, recent validations
Task Types	Label list, frequency, phase, manual validation trigger
Data	Training data stats, label distribution chart, export
Validation	Per-label validation details: pass/fail/reason
Skills	Auto-generated Skills list, view/delete
Router	Control which task types use cheap model offload

📊 Overview

See task counts, request volume, task type count, and validated count at a glance. Recent validation records with pass/fail reasons shown at the bottom.

🏷️ Task Types

All auto-classified task labels. View collection count, current phase (collecting → validating → validated → offloaded), pass/fail counts per type, with manual validation trigger.

📁 Data

Browse all collected training data. Filter by label or date, preview user messages and assistant responses. Pie chart shows label distribution. One-click export to JSONL or LLaMA-Factory format.

✅ Validation

Detailed validation records. Each entry shows: time, task type, cheap model used, result (pass/fail/not_mockable), and the primary model's specific judgment — e.g. "small model didn't maintain the Banana persona tone" or "small model correctly understood the task intent".

🧠 Skills

Auto-generated OpenClaw Skills. When the cheap model repeatedly fails at a task type, the primary model analyzes failure patterns and generates targeted SKILL.md files to help the cheap model perform better next time.

🔀 Router Control

Core operations panel. View each task type's routing status (🔵 primary / 🟢 cheap), pass rate, and one-click "Enable Offload" to switch to cheap model or "Disable Offload" to switch back. Supports manual switching at any phase.

Architecture

clawbuddy/
├── src/
│   ├── index.ts              # CLI (commander)
│   ├── proxy/
│   │   ├── server.ts         # Hono HTTP server (:18790)
│   │   ├── router.ts         # Route decision (primary / cheap)
│   │   ├── filter.ts         # Request filter (task / system-level)
│   │   └── token-cache.ts    # OAuth token cache (for background tasks)
│   ├── config/
│   │   ├── init.ts           # init / restore / config management
│   │   └── types.ts          # Type definitions
│   ├── store/
│   │   └── db.ts             # SQLite (better-sqlite3)
│   ├── collector/
│   │   ├── collector.ts      # Request collection + session chaining
│   │   ├── labeler.ts        # Cheap model labeling + merging
│   │   └── exporter.ts       # JSONL / LLaMA-Factory export
│   ├── validator/
│   │   ├── sandbox.ts        # Mock environment management
│   │   ├── mock-gen.ts       # Primary model generates setup scripts
│   │   ├── runner.ts         # Cheap model runs in sandbox
│   │   ├── judge.ts          # Primary model judges results
│   │   ├── scheduler.ts      # Async validation scheduler
│   │   ├── primary-call.ts   # Shared primary model caller (with auth)
│   │   └── trace-compress.ts # Task trace compression (truncate system prompts)
│   ├── skills/
│   │   ├── analyzer.ts       # Failure pattern analysis
│   │   ├── generator.ts      # SKILL.md generation
│   │   └── evidence.ts       # Evidence pool + pattern detection
│   └── dashboard/
│       ├── api.ts            # REST API
│       └── static/index.ts   # Single-page Dashboard
├── docs/
│   └── screenshots/          # Dashboard screenshots

Configuration

ClawBuddy config file is at ~/.clawbuddy/config.json:

{
  "mode": "mode2",                    // mode1 | mode2
  "proxy": {
    "port": 18790,                    // proxy port
    "dashboard_port": 18791           // dashboard port
  },
  "primary_api": {                    // Primary API (auto-read from OpenClaw)
    "provider": "zai",
    "base_url": "https://open.bigmodel.cn/api/coding/paas/v4",
    "api_key": "...",
    "model": "glm-5-turbo"
  },
  "cheap_api": {                      // Cheap small model
    "provider": "ollama",
    "base_url": "http://localhost:11434/v1",
    "api_key": "ollama",
    "model": "qwen3.5:9b"
  },
  "validation": {
    "min_tasks_to_start": 3,          // Min tasks to start validation
    "certify_after_n_passes": 5,      // Certify after N passes
    "schedule": "idle"                // idle | manual
  },
  "skills": {
    "auto_generate": true,            // Auto-generate Skills
    "trigger_after_n_failures": 3,    // Trigger after N failures
    "output_dir": "~/.openclaw/workspace/skills/"
  }
}

Cheap Model Options

ClawBuddy is compatible with any OpenAI-compatible API:

Provider	Example	Notes
Ollama	`--cheap-provider ollama --cheap-model qwen3.5:9b`	Local, zero cost
Ollama (large)	`--cheap-provider ollama --cheap-model qwen3.5:27b`	Local, better quality
vLLM	`--cheap-base-url http://localhost:8000/v1`	Self-hosted inference
Zhipu AI	`--cheap-base-url https://open.bigmodel.cn/api/...`	Cheap Chinese API
Gemini Flash	`--cheap-base-url ...`	Google's cheap model

Data Format

Exported JSONL format:

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}], "metadata": {"task_label": "poetry writing", "session_id": "abc123", "model": "glm-5-turbo", "timestamp": "2026-03-17T10:00:00Z"}}

Design Principles

Zero intrusion on OpenClaw: Only modifies the provider's baseUrl in config, no OpenClaw source code changes
Additive only: Generated Skills use clawbuddy- prefix, separate directories, never touch existing Skills
Continuous iteration, no permanent bans: Failed task labels pause and wait for new evidence, never permanently abandoned
User-controlled: All offload decisions confirmed by user, Dashboard allows viewing, editing, and rollback at any time
Data safety: All data stored in local SQLite, nothing uploaded

CLI Reference

Command	Description
`clawbuddy start`	One-click start (auto-backup, inject proxy, launch services)
`clawbuddy stop`	Stop and restore OpenClaw config
`clawbuddy status`	View status
`clawbuddy config show`	Show current config
`clawbuddy config set <key> <value>`	Edit config (supports dot notation)
`clawbuddy offload enable <label>`	Enable offload for a task type
`clawbuddy offload disable <label>`	Disable offload
`clawbuddy offload list`	List offloaded task types
`clawbuddy export`	Export training data
`clawbuddy validate <label>`	Manually trigger validation

Tech Stack

Language: TypeScript (ESM)
HTTP Server: Hono + @hono/node-server
Storage: SQLite (better-sqlite3)
CLI: Commander.js
Dashboard: HTML + Vanilla JS + Chart.js
Bundler: tsdown

License

MIT

ClawBuddy 🦞 中文文档

OpenClaw 的降本 buddy。让你的 AI 助手越用越便宜。

ClawBuddy 是 OpenClaw 的成本优化插件。它以透明代理的方式接管 OpenClaw 的 API 调用，在后台收集数据、验证小模型能力、自动生成 Skill——让你逐步把昂贵的大模型调用迁移到便宜的小模型上，而不牺牲任务完成质量。

工作原理

用户 → OpenClaw → ClawBuddy proxy (:18790) → 主 API (glm-5-turbo)
                         ↓
                ┌──────────────────────┐
                │      后台（异步）      │
                │                      │
                │  Mode 1: 数据收集     │
                │  Mode 2: mock 验证   │ ← cheap model (qwen3.5:9b) + skills
                └──────────────────────┘

ClawBuddy 在本地启动一个 OpenAI-compatible proxy server，修改 OpenClaw 配置将 API 请求指向它。对 OpenClaw 完全透明——它只看到一个普通的 model provider。

两种模式

📦 Mode 1: Data Distiller

纯数据收集，零侵入。

所有请求照常转发给大模型，用户体验零变化
后台自动收集每个 request 的完整 input/output
将同一 session 的多次请求串联成完整的 task 链路
用小模型给每个 task 自动打标签（如「文件整理」「代码重构」）
导出为 JSONL / LLaMA-Factory 格式，拿去 fine-tune 你的专属小模型

适合：有 GPU 资源，想通过训练从根本上提升小模型能力的用户。

🧪 Mode 2: Mock Validator + Skill Generator（推荐）

沙盒验证 + 自动生成 Skill，training-free 降本。

核心循环：

日常使用大模型 (glm-5-turbo) 完成任务
    ↓ 后台收集 task 链路
在 mock 沙盒中让小模型 (qwen3.5:9b) 执行同样的任务
    ↓ 大模型判定结果
pass → 积累通过记录
fail → 分析失败原因 → 自动生成 OpenClaw Skill → 带 Skill 重新验证
    ↓ 某类任务累计 N 个验证通过
Dashboard 通知 →「诗歌创作」已验证，建议切换到 cheap model
    ↓ 用户一键启用 offload

大模型只在三个低频点被调用：生成 mock 环境、判卷、生成 Skill。日常打标签等操作全部由小模型完成。

适合：没有 GPU 或不想训练模型，希望通过 prompt engineering 提升小模型的用户。

快速开始

安装

git clone https://github.com/haochen2115/clawbuddy.git
cd clawbuddy
npm install
npm run build
npm link  # 全局注册 clawbuddy 命令

使用

一条命令启动，一条命令停止：

# 启动（自动备份 OpenClaw 配置、注入 proxy、启动服务）
clawbuddy start

# 停止（Ctrl+C 或另一个终端运行）
clawbuddy stop

就这样。启动后正常使用 OpenClaw，一切照旧。打开 http://localhost:18791 查看 Dashboard。

start 会自动完成：

备份 OpenClaw 配置（首次启动时）
注入 proxy provider 到 OpenClaw（OpenClaw 热加载，无需重启）
启动 proxy server + dashboard

stop（或 Ctrl+C）会自动完成：

停止所有服务
精准还原 OpenClaw 配置（只移除 proxy 相关字段，不覆盖其他修改）

自定义配置

默认配置就能用，但你也可以通过启动参数调整：

# 指定模式（默认 mode2）
clawbuddy start --mode mode1

# 指定 cheap model
clawbuddy start --cheap-provider ollama --cheap-model qwen3.5:9b

# 或者修改配置文件
clawbuddy config set mode mode1
clawbuddy config set cheap_api.model qwen3.5:9b
clawbuddy config show

导出训练数据

# 导出全部
clawbuddy export --format jsonl --output ./training_data.jsonl

# 按任务类型筛选
clawbuddy export --label "文件整理" --output ./file_org.jsonl

# 导出为 LLaMA-Factory 格式
clawbuddy export --format llamafactory --output ./dataset/

管理 offload（Mode 2）

# 查看状态
clawbuddy status

# 启用 / 停用 offload
clawbuddy offload enable "诗歌创作"
clawbuddy offload disable "诗歌创作"
clawbuddy offload list

Dashboard

本地 Web Dashboard http://localhost:18791，包含 6 个面板：

面板	功能
概览	模式、任务总数、最近验证记录
任务类型	标签列表、频次、阶段、手动触发验证
数据	训练数据统计、标签分布图表、导出
验证	每个标签的验证详情：通过/失败/原因
Skills	自动生成的 Skills 列表，可查看/删除
路由	控制哪些任务类型启用 offload

📊 概览

一目了然地查看任务总数、请求量、任务类型数和已验证数量。底部展示最近的验证记录，包括 pass/fail 原因。

🏷️ 任务类型

所有被自动分类的任务标签一览。可以看到每种任务的收集数量、当前阶段（collecting → validating → validated → offloaded）、验证通过/失败次数，并支持手动触发验证。

📁 数据

浏览所有收集到的训练数据。支持按标签、日期筛选，可直接预览用户消息和助手回复。底部的饼图展示标签分布。支持一键导出为 JSONL 或 LLaMA-Factory 格式。

✅ 验证

详细的验证记录。每条记录展示：时间、任务类型、使用的 cheap model、验证结果（pass/fail/not_mockable），以及大模型给出的具体评判原因——比如"小模型未保持 Banana 人设的语气"或"小模型正确理解了任务意图"。

🧠 Skills

ClawBuddy 自动生成的 OpenClaw Skills 列表。当小模型在某类任务上反复失败时，大模型会分析失败模式并生成针对性的 SKILL.md，帮助小模型在下次执行时表现更好。

🔀 路由控制

核心操作面板。查看每种任务当前的路由状态（🔵 primary / 🟢 cheap），验证通过率，并可一键「启用 Offload」将任务切换到 cheap model，或「停用 Offload」切回大模型。支持任意阶段手动切换。

配置

ClawBuddy 配置文件位于 ~/.clawbuddy/config.json：

{
  "mode": "mode2",                    // mode1 | mode2
  "proxy": {
    "port": 18790,                    // proxy 端口
    "dashboard_port": 18791           // dashboard 端口
  },
  "primary_api": {                    // 主 API（自动从 OpenClaw 读取）
    "provider": "zai",
    "base_url": "https://open.bigmodel.cn/api/coding/paas/v4",
    "api_key": "...",
    "model": "glm-5-turbo"
  },
  "cheap_api": {                      // 便宜的小模型
    "provider": "ollama",
    "base_url": "http://localhost:11434/v1",
    "api_key": "ollama",
    "model": "qwen3.5:9b"
  },
  "validation": {
    "min_tasks_to_start": 3,          // 开始验证的最少 task 数
    "certify_after_n_passes": 5,      // 通过 N 次后认证
    "schedule": "idle"                // idle | manual
  },
  "skills": {
    "auto_generate": true,            // 自动生成 Skill
    "trigger_after_n_failures": 3,    // 失败 N 次后触发
    "output_dir": "~/.openclaw/workspace/skills/"
  }
}

Cheap Model 选择

ClawBuddy 兼容任何 OpenAI-compatible API 的模型：

Provider	配置示例	说明
Ollama	`--cheap-provider ollama --cheap-model qwen3.5:9b`	本地运行，零成本
Ollama (大)	`--cheap-provider ollama --cheap-model qwen3.5:27b`	本地运行，效果更好
vLLM	`--cheap-base-url http://localhost:8000/v1`	自建推理服务
智谱 AI	`--cheap-base-url https://open.bigmodel.cn/api/...`	国产便宜 API
Gemini Flash	`--cheap-base-url ...`	Google 便宜模型

数据格式

导出的 JSONL 格式：

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}], "metadata": {"task_label": "诗歌创作", "session_id": "abc123", "model": "glm-5-turbo", "timestamp": "2026-03-17T10:00:00Z"}}

设计原则

对 OpenClaw 零侵入：只改配置中的 baseUrl，不修改任何 OpenClaw 源码
只新增不修改：生成的 Skill 以 clawbuddy- 前缀命名，独立目录，不碰已有 Skill
持续迭代，不判死刑：task_label 验证失败后暂停等待新证据，不永久放弃
用户可控：所有 offload 决策由用户确认，Dashboard 上可随时查看、编辑、回退
数据安全：所有数据存储在本地 SQLite，不上传任何内容

CLI 命令一览

命令	说明
`clawbuddy start`	一键启动（自动备份、注入 proxy、启服务）
`clawbuddy stop`	停止并还原 OpenClaw 配置
`clawbuddy status`	查看状态
`clawbuddy config show`	查看当前配置
`clawbuddy config set <key> <value>`	修改配置（支持 dot notation）
`clawbuddy offload enable <label>`	启用 offload
`clawbuddy offload disable <label>`	停用 offload
`clawbuddy offload list`	查看已 offload 的任务
`clawbuddy export`	导出训练数据
`clawbuddy validate <label>`	手动触发验证

技术栈

语言：TypeScript (ESM)
HTTP Server：Hono + @hono/node-server
存储：SQLite (better-sqlite3)
CLI：Commander.js
Dashboard：HTML + Vanilla JS + Chart.js
构建：tsdown

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
docs/screenshots		docs/screenshots
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json
tsdown.config.ts		tsdown.config.ts

Folders and files

Latest commit

History

Repository files navigation

ClawBuddy 🦞

How It Works

Two Modes

📦 Mode 1: Data Distiller

🧪 Mode 2: Mock Validator + Skill Generator (Recommended)

Quick Start

Installation

Usage

Configuration

Export Training Data

Manage Offload (Mode 2)

Dashboard

📊 Overview

🏷️ Task Types

📁 Data

✅ Validation

🧠 Skills

🔀 Router Control

Architecture

Configuration

Cheap Model Options

Data Format

Design Principles

CLI Reference

Tech Stack

License

ClawBuddy 🦞 中文文档

工作原理

两种模式

📦 Mode 1: Data Distiller

🧪 Mode 2: Mock Validator + Skill Generator（推荐）

快速开始

安装

使用

自定义配置

导出训练数据

管理 offload（Mode 2）

Dashboard

📊 概览

🏷️ 任务类型

📁 数据

✅ 验证

🧠 Skills

🔀 路由控制

配置

Cheap Model 选择

数据格式

设计原则

CLI 命令一览

技术栈

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages