Agent Toolbelt
marras0914/agent-toolbelt
Análise de ações sintetizada por IA + 20 ferramentas utilitárias para agentes
Ver no GitHubConjunto de ferramentas especializadas para agentes de IA, com 7 ferramentas de análise de ações (dados financeiros ao vivo + síntese com IA) e 20 utilitários para infraestrutura de agentes. API em produção com modelo de pagamento por uso (1.000 chamadas/mês gratuitas).
Instalação
npx convez-ai marras0914/agent-toolbelt1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisa de ações e teses de investimento automatizadas
- Análise de earnings e sinais de insiders
- Comparação de valuation entre múltiplas empresas
- Extração de metadados, conversão de formatos e normalização de dados
- Automação de tarefas comuns em workflows de agentes
Veredito

Toolbelt prático e bem desenhado para agentes que trabalham com finanças e processamento de dados. As ferramentas de stock research entregam análise sintetizada (não raw data), economizando tokens; os utilitários cobrem casos reais. Pricing justo e API clara, embora dependa de fontes terceiras (Polygon, Finnhub, FMP).
README
A tradução para o português deste README ainda está na fila.
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.
Stock research tools (7)
LLM-synthesized analysis on top of live financial data from Polygon.io, Finnhub, and Financial Modeling Prep.
| Tool | What it does | Price |
|---|---|---|
stock-thesis | Full Motley Fool-style investment thesis: verdict + thesis paragraphs + strengths + risks + valuation read | $0.05/call |
earnings-analysis | EPS beat/miss history, revenue trend, long-term earnings consistency read, upcoming earnings date | $0.05/call |
insider-signal | Form 4 interpretation — distinguishes meaningful open-market purchases from routine sales/awards. Signal strength + confidence | $0.05/call |
valuation-snapshot | P/E, P/S, EV/EBITDA, FCF yield, ROE, margins → cheap/fair/expensive verdict + specific buy zone | $0.05/call |
bear-vs-bull | Steelmanned 3-bull / 3-bear case with specific data, net verdict, key debate question | $0.05/call |
compare-stocks | Head-to-head comparison of 2-3 tickers. Winner + per-ticker strengths/concerns + ifYouValue map (growth / value / quality) | $0.05/call |
moat-analysis | Buffett-style competitive moat assessment (brand / switching costs / network / scale / IP / cost). Wide/narrow/none + durability | $0.05/call |
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.
| Tool | What it does | Price |
|---|---|---|
text-extractor | Extract emails, URLs, phones, dates, currencies, addresses, names from text | $0.0005/call |
token-counter | Count tokens across 15 LLM models with cost estimates | $0.0001/call |
schema-generator | JSON Schema / TypeScript / Zod validator from plain English | $0.001/call |
csv-to-json | CSV to typed JSON with auto delimiter and type casting | $0.0005/call |
markdown-converter | HTML ↔ Markdown conversion | $0.0005/call |
url-metadata | Title, OG tags, favicon, author from any URL | $0.001/call |
web-summarizer | Fetch + summarize a URL with key points | $0.02/call |
regex-builder | Natural language → regex with JS/Python/TS snippets | $0.0005/call |
cron-builder | Schedule description → cron expression with next-run preview | $0.0005/call |
address-normalizer | US address → USPS format with component parsing | $0.0005/call |
color-palette | Color palettes with WCAG scores and CSS vars | $0.0005/call |
brand-kit | Full brand kit — colors, typography, CSS/Tailwind tokens | $0.001/call |
image-metadata-stripper | Strip EXIF/GPS/IPTC/XMP metadata for privacy | $0.001/call |
meeting-action-items | Action items, decisions, summary from meeting notes | $0.05/call |
prompt-optimizer | Score and rewrite LLM prompts | $0.05/call |
document-comparator | Semantic diff between two document versions | $0.05/call |
contract-clause-extractor | Key clauses + risk flags from contracts | $0.10/call |
api-response-mocker | Realistic mock data from a JSON Schema | $0.0005/call |
dependency-auditor | CVE scan for npm/PyPI packages via OSV database | $0.005/call |
context-window-packer | Pack content into a token budget for LLM context | $0.001/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
| Tier | Price | Monthly calls | Rate limit |
|---|---|---|---|
| Free | $0/mo | 1,000 | 10/min |
| PAYG | prepaid credits | unlimited | 60/min |
| Starter | $29/mo | 50,000 | 60/min |
| Pro | $99/mo | 500,000 | 300/min |
| Enterprise | Custom | 5,000,000 | 1,000/min |
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