terahood-agent
Enables agents to purchase and use compute resources on Robinhood Chain via x402 payments. Provides tools for thinking, renting GPUs, checking balances, and viewing the compute menu.
README
<div align="center">
⚡ TERAHOOD Agent
Your agent can't have a credit card. Give it a wallet instead.
An autonomous agent that buys its own compute on Robinhood Chain. No account. No API key. No card. The wallet is the login — and the burn is the receipt.
</div>
The problem
An AI agent with a wallet cannot buy inference from a closed provider.
Not "prefers not to" — cannot. No credit card. No KYC. No email inbox. No 2FA device. The billing rail that makes closed models convenient for humans makes them unreachable for software.
Meanwhile the buyer class is already here:
| ~69,000 | active agents paying onchain via x402 |
| ~165M | agent transactions processed |
| ~$50M | cumulative volume |
| 0 | of them hold a credit card |
<sub>x402, by April 2026</sub>
They have money and nowhere to spend it. This is the door.
Install
npm install @terahood/agent
30 seconds
import { Agent } from "@terahood/agent";
const agent = new Agent({
privateKey: process.env.AGENT_PRIVATE_KEY,
endpoint: process.env.FORGE_ENDPOINT,
});
const r = await agent.think("Is compute a real-world asset?");
console.log(r.answer);
console.log(r.burnRef); // proof the job ran, on-chain
That's the whole integration. No signup step, because there isn't one.
What it does when you're not looking
await agent.loop(() => queue.next(), {
onResult: (r) => console.log(r.burnRef, r.answer),
});
It checks its balance. If the next job would leave it short, it mints more grams itself and keeps going. No human approves anything.
$ agent.tick()
balance 0g · need 0.15g
$ bars.mint(1g)
minted 1g · signed by 0xff6e1ea7…
$ forge.think("should I reserve H100 capacity?")
routing → Forge 70B · Llama 3.3 70B · 0.15g
HTTP 402 payment required · 0.15g
paid · 0x59ea…7e6e2 · retrying with X-PAYMENT
inference · 583 tokens
│ Reserve. New hardware is 36–52 weeks out and on-demand
│ runs 2–3x at peak. 8x H100 for 40h = 448g.
receipt BRN-46BA18A5 · 0xf1b5…974655 → ledger
How a payment actually works
sequenceDiagram
participant A as Agent
participant F as Forge
participant C as Robinhood Chain
A->>F: POST /x402/forge
F-->>A: 402 Payment Required<br/>{ asset, amount, nonce }
A->>A: sign terms with wallet key
A->>F: POST /x402/forge<br/>X-PAYMENT: …
F->>C: burn(grams, deliveryHash)
C-->>F: burnRef
F-->>A: 200 + answer<br/>X-PAYMENT-RESPONSE: { burnRef }
The signature is the authorization. There is no account to create, because the wallet already is one.
Compute Bars
The asset the agent spends. Prepaid GPU-hours, minted on-chain, denominated like gold ingots.
1 gram ≈ $1 of compute. Transferable — hold them, trade them, or burn them.
| Model | Cost | Class |
|---|---|---|
forge-8b |
0.05g | Fast answer |
mixtral |
0.10g | Fast answer |
coder-33b |
0.12g | Code |
forge-70b |
0.15g | Deep reasoning |
qwen-72b |
0.15g | Deep reasoning |
deepseek-v3 |
0.20g | Deep reasoning |
One bar, two ways to spend it: finished answers (above) or raw GPU hours (0.50g/hr RTX 4090 → 1.40g/hr H100 SXM).
Three doors
However your agent already talks, there's a way in.
<table> <tr><td width="33%">
OpenAI-compatible
Point any existing SDK at the Forge. The migration cost is a base URL.
</td><td width="33%">
x402
402 → pay → retry with receipt. The protocol 69,000 agents already speak.
</td><td width="33%">
MCP
Plug the Forge in as a tool. It discovers the menu and buys.
</td></tr> </table>
# MCP
claude mcp add foundry -- node foundry-mcp.js
# tools: forge_think · forge_menu · forge_rent_gpu · forge_ledger · forge_key_balance
Every job proves itself
Ask any AI provider to prove a specific inference ran — on the model you paid for, at the time they billed you. They can't. You get a dashboard and their word.
Every burn here writes a burn ref and a delivery hash of the request and response to a public ledger on Robinhood Chain.
const r = await agent.think("…");
r.burnRef; // BRN-46BA18A5
r.deliveryHash; // 0xf1b5…974655
Proof-of-reserve, except the reserve delivers itself.
API
new Agent({ privateKey, endpoint, rpcUrl?, minBalance? })
| Method | Returns |
|---|---|
agent.address |
the agent's wallet address |
await agent.balance() |
grams held |
await agent.mint(grams) |
tx hash |
await agent.ensureFunded(need) |
tops up if short, returns balance |
await agent.think(prompt, { model }) |
{ answer, model, grams, burnRef, deliveryHash } |
await agent.menu() |
available models + gram prices |
await agent.loop(nextPrompt, opts) |
runs forever |
Config
cp .env.example .env
| Var | What |
|---|---|
AGENT_PRIVATE_KEY |
the agent's wallet — use a burner, this is an autonomous spender |
FORGE_ENDPOINT |
Forge base URL |
RPC_URL |
optional; defaults to the public Robinhood Chain RPC (rate limited) |
Network — Robinhood Chain, chainId 4663, ETH gas, ~100ms blocks, fully EVM.
ComputeBar — 0xb32732c5b27b14cd7e0911fdb1f5a03b1cf86715
⚠️ Fund the agent's key with only what you're willing to let it spend. It mints without asking — that's the entire point.
Status
Early. The SDK and the x402 flow are implemented against the endpoint above; expect the interface to move before 1.0. Issues and PRs welcome.
Why this exists
Robinhood is bringing real-world assets on-chain, and compute is the cleanest one on the list — real, scarce, dollar-priced, and 100% digital. No vault, no courier, no custodian. It delivers itself and signs its own receipt.
Agents were handed wallets. They needed something to buy.
<div align="center">
terahood.xyz · @terahood_xyz · Proof Ledger
Compute, made tangible.
</div>
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。