How do I give an OpenAI Agents SDK agent access to fund data?
Add a function tool that calls the fund API by ISIN, or attach the hosted MCP server, so an OpenAI Agents SDK agent answers fund questions with current figures. Code for both.Published September 12, 2026 · Updated October 13, 2026 · by FundFacts APIShort answer
Either wrap one API call in a @function_tool (Python) or tool() (TypeScript) that returns the fields you want the agent to see, or attach the FundFacts MCP server (https://fundfactsapi.com/api/mcp) as a hosted MCP tool with an Authorization header. The agent then calls get_fund, search_funds, compare_funds, analyze_portfolio or fund_overlap as needed and answers with the current published figures and their as-of date.
Function tool (Python)
python# pip install git+https://github.com/fundfactsapi/fundfactsapi-python openai-agentsfrom agents import Agent, Runner, function_toolfrom fundfacts import FundFactsff = FundFacts() # reads FUNDFACTS_API_KEY@function_tooldef fund_facts(isin: str) -> dict:"""Current factsheet for a fund or ETF ISIN: name, TER, 1-7 risk, category, top holdings, sector/country split, returns, as-of date."""f = ff.get_fund(isin)d = f["data"]return {"name": f["name"], "ter": d["headlineMetrics"]["ter"], "risk": d["riskRating"], "category": d["profile"]["category"],"top_holdings": d["topHoldings"][:10], "sectors": d["sector"][:8], "countries": d["geography"][:8],"calendar_returns": d["calendarReturns"], "as_of": d["dataAsOf"]}agent = Agent(name="Fund analyst", instructions="Answer with figures from the tool and always state as_of. Empty fields are 'not disclosed'.", tools=[fund_facts])print(Runner.run_sync(agent, "Compare the cost and sector mix of IE00B4L5Y983 and IE00B3RBWM25").final_output)
Hosted MCP tool
The Agents SDK and the Responses API can call a remote MCP server on your behalf; no wrapper code:
pythonfrom agents import Agent, HostedMCPToolagent = Agent(name="Fund analyst",tools=[HostedMCPTool(tool_config={"type": "mcp","server_label": "fundfacts","server_url": "https://fundfactsapi.com/api/mcp","headers": {"Authorization": "Bearer ffk_..."},"require_approval": "never",})],)
typescript// npm install @openai/agentsimport { Agent, hostedMcpTool, run } from "@openai/agents";const agent = new Agent({name: "Fund analyst",instructions: "Quote dataAsOf with every figure.",tools: [hostedMcpTool({ serverLabel: "fundfacts", serverUrl: "https://fundfactsapi.com/api/mcp", headers: { Authorization: `Bearer ${process.env.FUNDFACTS_API_KEY}` } })],});const result = await run(agent, "How much do IE00B4L5Y983 and IE00B3RBWM25 overlap?");console.log(result.finalOutput);
The MCP server exposes seven tools, each counted like the REST endpoint it wraps: get_fund (ISIN → the full factsheet payload), search_funds (name → ISINs, free), compare_funds (side-by-side table for several ISINs), analyze_portfolio (weighted look-through of a list of positions, Starter and above), fund_overlap (shared holdings between funds, Starter and above), get_scpi (French SCPI data by name or ISIN) and search_scpi (SCPI name → matches, free), plus a tour prompt that explains the service on the example fund IE00B4L5Y983, which is not counted. Every tool only reads data. Authentication is either your FundFacts account (OAuth 2.1 with PKCE, the sign-in the claude.ai, Claude Desktop and ChatGPT connectors run) or the same bearer API key as the REST API, sent as an Authorization header by editors, CLIs and frameworks. The MCP server guide has the full tool reference, setup for every client and troubleshooting.
Production notes
Give function tools a 300-second timeout for cold ISINs, cache by ISIN until expiresAt, and tell the agent that a 404 means "not a covered fund". Tool calls cost one request per fund found, the same as REST; misses are free.
Verify it yourself
The demo endpoint returns the live payload for a fund that is already in the store, without a key. Everything on this page can be checked against it.bashcurl -s https://fundfactsapi.com/api/v1/demo/funds/IE00B4L5Y983 | jq '{name, asOf: .data.dataAsOf, ter: .data.headlineMetrics.ter, risk: .data.riskRating, top: .data.topHoldings[:3]}'