Agent Toolbelt

Stock research tools for AI agents. Live financial data + Claude-synthesized analysis, served as 7 focused tools — not raw OHLCV. Plus 20 utility tools for the rest of an agent's work.

Production API: https://www.agenttoolbelt.live

Quickstart

# Get a free API key (1,000 calls/month, no credit card)
curl -X POST 'https://www.agenttoolbelt.live/api/clients/register' \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com"}'

# Generate a Motley Fool-style investment thesis for any ticker
curl -X POST https://www.agenttoolbelt.live/api/tools/stock-thesis \
  -H "Authorization: Bearer atb_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"ticker": "NVDA", "timeHorizon": "3-5 years"}'

Returns: bullish/neutral/bearish verdict, thesis paragraphs, key strengths, key risks, valuation read, insider read, analyst consensus read, and what to watch in the next earnings report.

Related MCP server: agent-utils-mcp

Stock research tools (7)

LLM-synthesized analysis on top of live financial data from Polygon.io, Finnhub, and Financial Modeling Prep.

Every stock tool returns a dataSources block with fetchedAt + per-source success flags so you know exactly what data backed the analysis.

Utility tools (20)

Common agent infrastructure. Rule-based tools billed at $0.0001–$0.001/call; LLM-powered tools at $0.005–$0.10/call.

Do you know what your agent is actually calling?

You wired up the tools. Once the agent is running, though, you have no visibility into which ones it calls or what it passes to them. I also built Cordon for this. It sits in front of any MCP server and logs every call. You can write rules to block the ones you don't want touching production.

getcordon.com — npx @getcordon/cli init

npm SDK + LangChain

npm install agent-toolbelt

Typed client

import { AgentToolbelt } from "agent-toolbelt";

const client = new AgentToolbelt({ apiKey: process.env.AGENT_TOOLBELT_KEY! });

// Stock research
const thesis = await client.stockThesis({ ticker: "NVDA", timeHorizon: "3-5 years" });
const moat = await client.moatAnalysis({ ticker: "AAPL" });
const compare = await client.compareStocks({ tickers: ["NVDA", "AMD"] });

// Utility
const tokens = await client.tokenCounter({ text: myDocument });
const contacts = await client.textExtractor({
  text: emailBody,
  extractors: ["emails", "phone_numbers", "addresses"],
});

LangChain integration

import { AgentToolbelt } from "agent-toolbelt";
import { createLangChainTools } from "agent-toolbelt/langchain";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { ChatOpenAI } from "@langchain/openai";

const client = new AgentToolbelt({ apiKey: process.env.AGENT_TOOLBELT_KEY! });
const tools = createLangChainTools(client); // 27 ready-to-use DynamicStructuredTools

const agent = createReactAgent({
  llm: new ChatOpenAI({ model: "gpt-4o" }),
  tools,
});

Claude MCP

Use the stock research tools (and the rest of the toolbelt) directly inside Claude Desktop or Claude Code via the agent-toolbelt-mcp package.

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "agent-toolbelt": {
      "command": "npx",
      "args": ["-y", "agent-toolbelt-mcp"],
      "env": {
        "AGENT_TOOLBELT_KEY": "atb_your_key_here"
      }
    }
  }
}

Claude Code — one command:

claude mcp add agent-toolbelt -e AGENT_TOOLBELT_KEY=atb_your_key_here -- npx -y agent-toolbelt-mcp

Once installed, ask Claude things like "Give me a full analysis of NVDA — thesis, earnings quality, insider activity, and whether it's cheap right now" and it'll call the tools in parallel.

Discover tools programmatically

Agents can auto-discover all 27 tools at runtime:

curl https://www.agenttoolbelt.live/api/tools/catalog

Pricing

Integrations

• npm SDK — npm install agent-toolbelt — typed client + LangChain tools

• MCP — npx -y agent-toolbelt-mcp — works with Claude Desktop and Claude Code

• OpenAI GPT Actions — OpenAPI spec at /openapi/openapi-gpt-actions.json

• RapidAPI — listed on the RapidAPI marketplace

• Smithery, Glama, PulseMCP, MCP registry — discoverable in MCP directories

Going to production

When the agent moves out of dev, a new set of questions shows up. What did it call last night? What arguments did it pass? Who approved the destructive one?

Cordon is an MCP gateway that sits in front of servers like this one. Point your client at Cordon instead of directly at Agent Toolbelt; Cordon forwards every call through and adds:

• A real-time audit log of every tool invocation (name, arguments, response, latency)

• Per-API-key policy: which tools each caller can use, under what conditions

• Slack-based human approvals for tool calls you've flagged as high-risk

From the agent's perspective nothing changes — same tools, same schemas. Free tier covers 1,000 events/month.

License

MIT