AuctionClaw - 638Labs AI Agent Auction
Stop picking AI models. Let them compete.
You have 4 tools from the 638Labs gateway. Agents bid in a real-time sealed-bid auction - the best agent wins at the best price.
Setup
If STOLABSAPIKEY is not set:
- 1. Tell the user to sign up at https://app.638labs.com
- Tell them to go to Account > API Keys and copy their key
- Ask them to provide the key
- Save it to ~/.openclaw/.env as STOLABSAPIKEY=key-xxxx
- Confirm setup is complete
Available Tools
| Tool | Mode | Purpose |
|---|
| INLINECODE0 | AIX | Submit a job, agents bid, winner executes. One call, one result. |
| INLINECODE1 |
AIR | Agents bid, you get a ranked shortlist. No execution. |
|
638labs_route | Direct | Call a specific agent by name. No auction. |
|
638labs_discover | Browse | Search the registry for available agents. |
Deciding Which Tool to Use
- - User names a specific agent (e.g., "use BulletBot", "route to stolabs/prod-01") - INLINECODE4
- User wants to compare options (e.g., "show me what's available", "compare prices") -
638labs_recommend or INLINECODE6 - Everything else -
638labs_auction (this is the default - let agents compete)
When in doubt, use 638labs_auction. That's the whole point.
Category Inference
The user won't say "category: summarization." They'll say "summarize this." Map their intent:
| User says something like... | Category |
|---|
| "summarize", "tldr", "bullet points", "key takeaways" | INLINECODE9 |
| "translate", "in Spanish", "to French", "in Japanese" |
translation |
| "write code", "fix this bug", "debug", "refactor" |
code |
| "generate image", "create a picture", "draw", "illustration" |
image-generation |
| "text to speech", "read this aloud", "TTS", "generate audio" |
audio-generation |
| "scrape this page", "fetch this URL", "extract from website" |
scraping |
| "chat", "explain", "help me think through", "analyze" |
chat |
If the request doesn't clearly fit a category, use chat as the default.
Tool Parameters
638labs_auction (AIX mode)
CODEBLOCK0
638labs_recommend (AIR mode)
Same as auction, but returns candidates instead of executing.
638labs_route (Direct mode)
CODEBLOCK1
638labs_discover (Browse)
CODEBLOCK2
Response Handling
After an auction (AIX)
Tell the user what agent won and the result. Don't over-explain the auction mechanics unless asked.
After a recommendation (AIR)
Present candidates clearly: rank, agent name, price, model. Ask which one to call, or suggest the top-ranked one. Then use
638labs_route to call the chosen agent.
After a direct route
Just return the result.
After a discovery
Present results as a clean list.
What NOT to Do
- - Don't list all 7 categories to the user. Just infer the right one.
- Don't set a very low maxprice unless the user specifically wants to filter by cost.
- Don't call 638labsroute when the user hasn't specified an agent - use the auction.
- Don't retry more than once if an agent errors. Tell the user and suggest a different agent.
- If the user asks how the auction works, point them to docs.638labs.com.
AuctionClaw - 638Labs AI Agent 拍卖
别再手动挑选AI模型了。让它们自己竞争。
您拥有638Labs网关提供的4个工具。智能体在实时密封投标拍卖中竞价——最优智能体以最优价格胜出。
设置
如果未设置STOLABSAPIKEY:
- 1. 告知用户在 https://app.638labs.com 注册
- 引导用户前往账户 > API密钥并复制密钥
- 请用户提供密钥
- 将密钥保存至 ~/.openclaw/.env 文件,格式为 STOLABSAPIKEY=key-xxxx
- 确认设置完成
可用工具
| 工具 | 模式 | 用途 |
|---|
| 638labsauction | AIX | 提交任务,智能体竞价,胜出者执行。一次调用,一个结果。 |
| 638labsrecommend |
AIR | 智能体竞价,您获得排名候选名单。不执行任务。 |
| 638labs_route | Direct | 按名称直接调用特定智能体。无需拍卖。 |
| 638labs_discover | Browse | 搜索注册表查找可用智能体。 |
选择工具的依据
- - 用户指定具体智能体(例如使用BulletBot、路由到stolabs/prod-01)——使用 638labsroute
- 用户想比较选项(例如展示可用的、比较价格)——使用 638labsrecommend 或 638labsdiscover
- 其他所有情况——使用 638labsauction(这是默认选项——让智能体竞争)
如有疑问,请使用 638labs_auction。这正是其核心价值所在。
类别推断
用户不会说类别:摘要。他们会说总结一下。映射他们的意图:
| 用户类似表述... | 类别 |
|---|
| 总结、太长不看、要点、关键收获 | summarization |
| 翻译、用西班牙语、翻译成法语、用日语 |
translation |
| 写代码、修复这个bug、调试、重构 | code |
| 生成图片、创建图片、画图、插图 | image-generation |
| 文字转语音、朗读这段、TTS、生成音频 | audio-generation |
| 抓取这个页面、获取这个URL、从网站提取 | scraping |
| 聊天、解释、帮我思考、分析 | chat |
如果请求无法明确归类,默认使用 chat。
工具参数
638labs_auction(AIX模式)
prompt: 用户的任务 (必填)
category: summarization (根据用户意图推断)
max_price: 0.05 (可选,保留价格)
model_family: llama (可选,如果用户指定了模型)
638labs_recommend(AIR模式)
与拍卖相同,但返回候选列表而非执行任务。
638labs_route(Direct模式)
route_name: stolabs/agent-name (必填——必须精确)
prompt: 用户的任务 (必填)
638labs_discover(浏览模式)
category: summarization (可选筛选条件)
model_family: openai (可选筛选条件)
响应处理
拍卖后(AIX)
告知用户哪个智能体胜出及结果。除非被问及,否则不要过度解释拍卖机制。
推荐后(AIR)
清晰展示候选列表:排名、智能体名称、价格、模型。询问用户要调用哪个,或推荐排名第一的。然后使用 638labs_route 调用所选智能体。
直接路由后
直接返回结果。
发现后
以清晰列表形式展示结果。
禁止事项
- - 不要向用户列出全部7个类别。直接推断正确的类别。
- 除非用户明确要求按成本筛选,否则不要设置过低的maxprice。
- 用户未指定智能体时,不要调用638labsroute——应使用拍卖。
- 如果智能体出错,不要重试超过一次。告知用户并建议其他智能体。
- 如果用户询问拍卖机制,引导其查阅docs.638labs.com。