Integrations — AgentBroker

// 01 / langchain

LangChain

AgentBrokerToolkit returns six tools — get_prices, place_order, get_balance, get_portfolio, search_tokens, get_candles — ready to drop into any LangGraph agent. Reads AGENTBROKER_API_KEY from env, no key in source.

pip install agentbroker langchain-anthropic langgraph

Show 10-line snippet

import os
from agentbroker.integrations.langchain import AgentBrokerToolkit
from langchain_anthropic import ChatAnthropic
from langgraph.prebuilt import create_react_agent
from agentbroker import AgentBroker

os.environ["AGENTBROKER_API_KEY"] = "ab_live_…"
llm = ChatAnthropic(model="claude-3-5-sonnet-20241022")
agent = create_react_agent(llm, AgentBrokerToolkit().get_tools())
print(agent.invoke({"messages": [("user", "Buy 0.05 SOL if it's under $150")]}))
await AgentBroker(api_key=os.environ["AGENTBROKER_API_KEY"]).trade(
    "SOL", "USDC", amount_sol=0.05, dry_run=True)

View on PyPI ↗ Source ↗

// Test wallet only · No sign-in · No real funds.

// 02 / crewai

CrewAI

Four plug-and-play tools — Price, Candles, Trade, Portfolio — for any CrewAI role. Ships with a ready-made two-agent analyst + trader crew (example_crew.py) that reads OHLCV, picks a setup, and executes the order.

pip install agentbroker crewai

Show 10-line snippet

import os
from crewai import Agent
from integrations.crewai.tools import (
    AgentBrokerPriceTool, AgentBrokerCandlesTool,
    AgentBrokerTradeTool, AgentBrokerPortfolioTool)
from integrations.crewai.example_crew import build_trading_crew

os.environ["AGENTBROKER_API_KEY"] = "ab_live_…"
trader = Agent(role="Crypto Trader",
               goal="Trade profitably on AgentBroker",
               tools=[AgentBrokerPriceTool(),
                      AgentBrokerTradeTool(api_key=os.environ["AGENTBROKER_API_KEY"])])
crew = build_trading_crew(api_key=os.environ["AGENTBROKER_API_KEY"],
                          trade_budget_usdc=50, dry_run=True)
_ = crew.kickoff()

Source ↗

// Test wallet only · No sign-in · No real funds.

// 03 / autogpt

AutoGPT

One drop-in file (agentbroker_block.py) registers four blocks: GetPrice, GetCandles, PlaceOrder, GetBalance. Copy it into AutoGPT Platform's backend/blocks/, restart, and the blocks are discoverable to every agent.

pip install autogpt-platform

Show 10-line snippet

import sys; sys.path.insert(0, "backend/blocks/")
from backend.blocks.agentbroker_block import (
    AgentBrokerGetPriceBlock,
    AgentBrokerTradeBlock,
)

price = AgentBrokerGetPriceBlock().run(
    AgentBrokerGetPriceBlock.Input(pair="SOL-USDC")
)
print("SOL", price["price_usd"])
trade = AgentBrokerTradeBlock().run(
    AgentBrokerTradeBlock.Input(pair="SOL-USDC", side="buy",
                                amount_usdc=10, dry_run=True)
)
print("order", trade["order_id"])

Source ↗

// Test wallet only · No sign-in · No real funds.

// 04 / claude-mcp

Claude MCP

One claude mcp add command wires the AgentBroker MCP server into Claude Code / Claude Desktop. Discover register_agent, get_prices, trade and the rest via the standard MCP tools/list handshake — no SD K to install, no schema to paste.

claude mcp add io.github.polsia/agentbroker --transport http https://agentbroker.polsia.app/mcp

Show 10-line snippet

# 1. Install via Claude Code (per Anthropic's MCP add syntax):
claude mcp add io.github.polsia/agentbroker \
  --transport http https://agentbroker.polsia.app/mcp

# 2. JSON-RPC probe — what Claude Desktop sends on connect:
curl -X POST https://agentbroker.polsia.app/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"get_prices","arguments":{}}}'

# 3. Or via the python-sdk after register_agent:
from agentbroker import AgentBroker
_ = await AgentBroker(api_key="ab_…").trade("SOL", "USDC", amount_sol=0.05, dry_run=True)

MCP source ↗

// Test wallet only · No sign-in · No real funds.

// under the hood

Every framework ships with the same six tools behind it. Whatever the stack, the agent gets live Jupiter prices, OHLCV candles, market and limit orders, and a portfolio view — all over REST, all keyed off a single API key.

AgentBroker is non-custodial by design. Agents supply their own Solana wallet, AgentBroker keeps no deposit, and every fill is confirmed on-chain. There is no platform-side key custody and no human-in-the-loop approval — the agent is the principal.

The canonical live path is exactly three AgentBroker requests: POST /v1/onboarding/first-trade/register, POST /v1/onboarding/first-trade/fund-and-prepare with a finalized mainnet USDC signature and trade intent, then local signing/broadcast followed by POST /v1/onboarding/first-trade/confirm. Python and AutoGen expose the same operations; neither stores keys, signs, or submits Solana transactions.

SOL-USDC, BTC-USDC, ETH-USDC, JUP-USDC, BONK-USDC and WIF-USDC are the default pairs the quote endpoints cover. Every other pair AgentBroker routes via Jupiter is reachable through search_tokens and get_candles — discoverable, not curated.

Connect AgentBroker to the framework your agent already speaks.