@hoodstack/agent-kit

Give AI agents safe, typed access to Robinhood Chain through HoodStack.

It ships two things from one small package:

• A framework-agnostic toolset — drop it into the Vercel AI SDK, LangChain, or a raw function-calling loop.
• An MCP server — expose the same tools to any Model Context Protocol client (Claude Desktop, Claude Code, and others) with one config block.

Every tool is a read or a simulation. Nothing here signs a message, moves funds, or submits a transaction, so an agent can plan and verify safely. (Signed, policy-bounded execution is on the roadmap — see Roadmap.)

npm install @hoodstack/agent-kit
# or: pnpm add @hoodstack/agent-kit

You need a HoodStack project API key (hs_test_… for testnet, hs_live_… for mainnet). Create one in the dashboard.

All of these run against the live HoodStack API — authenticated by your key, rate limited, and metered.

The package installs a hoodstack-mcp command. Point an MCP client at it and set your key.

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "hoodstack": {
      "command": "npx",
      "args": ["-y", "@hoodstack/agent-kit", "hoodstack-mcp"],
      "env": { "HOODSTACK_API_KEY": "hs_test_your_key" }
    }
  }
}

Claude Code:

claude mcp add hoodstack -e HOODSTACK_API_KEY=hs_test_your_key \
  -- npx -y @hoodstack/agent-kit hoodstack-mcp

Then ask your assistant things like "What's the current gas on Robinhood Chain?" or "Simulate sending 0.01 ETH to 0x… and tell me if it would succeed."

import { createHoodStackAgent } from "@hoodstack/agent-kit";

const agent = createHoodStackAgent({ apiKey: process.env.HOODSTACK_API_KEY! });

// Run a tool by name.
const gas = await agent.run("hoodstack_get_gas", {});
const sim = await agent.run("hoodstack_simulate_transaction", {
  to: "0x…",
  valueWei: "10000000000000000",
});

// Or inspect the definitions to wire into your own framework.
for (const tool of agent.tools) {
  console.log(tool.name, "-", tool.description);
}

Each tool is a plain object — { name, description, inputSchema, execute } — where inputSchema is a Zod raw shape. That adapts cleanly to most frameworks. For the Vercel AI SDK, for example:

import { tool } from "ai";
import { z } from "zod";
import { createHoodStackAgent } from "@hoodstack/agent-kit";

const hoodstack = createHoodStackAgent({ apiKey: process.env.HOODSTACK_API_KEY! });

const tools = Object.fromEntries(
  hoodstack.tools.map((t) => [
    t.name,
    tool({
      description: t.description,
      parameters: z.object(t.inputSchema),
      execute: (input) => hoodstack.run(t.name, input),
    }),
  ]),
);

• Non-custodial. HoodStack cannot move user funds, and this kit never asks it to — it only reads and simulates. There is no signing key anywhere in this package.
• Simulate before acting. hoodstack_simulate_transaction lets an agent check whether a call would succeed (and what it would cost) before anyone commits to it — the right default for autonomous behavior.
• One metered spine. Every call is authenticated by your key, rate limited, and metered by HoodStack, so an agent can't run away with your quota unseen.

• Signed execution. Submitting transactions from an agent's smart account — owned by the agent's key, bounded by server-side spend limits and allowlists, relayed by HoodStack — is proven on testnet and being wired into the platform. A hoodstack_send_transaction tool lands here once it's live end to end. It will stay non-custodial: the agent signs, the relayer only relays.
• Sessions. Scoped, expiring permissions (session keys) so an agent gets exactly the capability it needs, for exactly as long as it needs it.

Until those ship, this kit stays read- and simulate-only — honestly, that covers most of what a safe agent should do before it ever acts.

Apache-2.0. HoodStack is an independent project and is not affiliated with, endorsed by, or operated by Robinhood Markets, Inc.