How to build an AI agent: Finance Research Agent

How to build an AI agent: Finance Research Agent

Let's say you're creating a CLI agent that can plan investment strategies. You want your agent to be capable of answering questions about a company’s performance while also pulling in context from the web.

With the Dedalus SDK, this is straightforward. You can connect multiple MCP servers from our marketplace and route each request to the right tool. In this example, explanatory questions go to the Exa web-search server, while questions about a company’s financial standing go to a Yahoo Finance server.

GitHub repo: dedalus-labs/finance-research-agent

How we built this

The agent is built with the Dedalus runner and configured with mcp_servers=[...]. Dedalus acts as an MCP client: it connects to multiple MCP servers, discovers their tool schemas, and lets the model call them without having to customize integration code for each provider. It also hosts the servers, configuring MCP for you.

Dedalus’ streaming support lets the CLI to stream responses as the tools run. Unlike raw SSE streaming available in most SDKs, we provide semantic events:

with client.chat.completions.stream(
	model="claude-opus-4.5",
	messages=[...],
	response_format=MyPydanticModel,  # Structured output preserved through streaming
) as stream:
	for event in stream:
		if event.type == "content.delta":
			print(event.delta, end="")
		elif event.type == "content.done":
			obj = event.parsed  # Already a validated Pydantic instance

Semantic event streaming means state accumulation happens automatically, so follow-up questions like “compare that to X” can be answered without rebuilding state manually. It also stops models from getting bogged down through resource cleanup via context managers.

Other features of semantic events streaming include:

• Structured output parsing that works mid-stream
• Tool call arguments that stream incrementally with tool_calls.function.arguments.delta

Behind this CLI are two MCP servers from the Dedalus MCP marketplace, a central hub for tools your agent can use. A market-data server provides quotes, historical OHLCV, and basic company financials. A web research server lets the agent cite news, filings, and reports from the open internet as it makes recommendations.

The tool calls behind this agent enable complex questions with answers users can trust. For example:

• “Compare NVDA vs AMD over the last 6 months: total return, volatility, max drawdown. Show the numbers.”
• “Now explain the biggest 3 drivers for each using sources from the last 30 days. Include links.”
• “Pull the latest quarterly revenue and gross margin for both and summarize trend direction.”
• “Compare that to INTC over the same window and explain the divergence using sources.”

Try building something

This is just one example of what you can build with the Dedalus SDK and marketplace. MCP gives your agents a structured interface to tools and data, so they can answer richer questions with better context and interface the world to take actions.

If you want to build something like the Finance Research Agent, you can start using our SDK for free in just a few lines of code. You can also access the Finance Research Agent GitHub repo. If you are not sure where to start, check out our docs for more in-depth tutorials and join our developer community on Discord. Happy building!