AI Research Agents & Copilots

Everything in this listing does the reading itself. You hand it a ticker or a question, a model works through filings, prices, news or transcripts, and it comes back with a report. The job is the same throughout; the delivery form decides most of what follows. A framework is open-source code you run with your own model key and your own data keys: you pay per token, you can read every prompt, and nobody supports it. A copilot is a hosted service you subscribe to: someone else holds the corpus and the model, and what you can check is the citation.

• Investor personas scoring US stocks, replayed as a fund — AI Hedge Fund, MIT, on a paid Financial Datasets key.
• Analysts, a bull, a bear and risk agents debating one ticker — TradingAgents, Apache-2.0, keyless yfinance data by default.
• An analyst-style report built from statements and a DCF — FinRobot, Apache-2.0, OpenAI and FMP keys.
• A local research workspace with backtests and broker connectors — Vibe-Trading, MIT, about twenty model providers.
• AI search over filings, broker research and expert transcripts — AlphaSense, enterprise contracts only.
• Discontinued — Fintool, bought by Microsoft on 18 April 2026.

What decides the choice

Which model, and what a run costs. None of the frameworks ships a model. AI Hedge Fund takes nine models across six providers, TradingAgents and Vibe-Trading a dozen or more providers each, local Ollama included, and FinRobot's report generator is built on the OpenAI SDK. Nobody publishes a cost per run, so reason from the structure. TradingAgents at default depth is four analyst legs, two debate turns, three risk turns and three manager and trader calls, and its "Deep" preset makes the two debate turns ten. AI Hedge Fund caches every prompt, so a re-run over unchanged filings costs nothing. Vibe-Trading's own guidance is that small models fabricate answers instead of calling tools: the cheap model is not the cheap run.

What it reads. AI Hedge Fund reads Financial Datasets and nothing else, and its personas see fundamentals alone — no prices, no news. FinRobot reads Financial Modeling Prep, whose free plan allows 250 calls a day. TradingAgents defaults to yfinance, and Vibe-Trading to a fallback chain of 27 sources, most of them keyless public endpoints nobody promised to keep. AlphaSense's contract buys a corpus no framework can reach: sell-side research and Tegus expert transcripts among 500 million-plus documents.

What comes out. AI Hedge Fund: a conviction between −1 and +1 and a written thesis per persona, blended by deterministic code into target weights and replayed in a backtest with no trading costs. TradingAgents: a tree of Markdown reports ending on a rating from Buy to Sell, with an entry, a stop, sizing and an optional price target. FinRobot: an HTML or PDF report whose cover rating and target are the latest sell-side figures reprinted from FMP, with the model writing the thesis around them. Vibe-Trading: a run card with metrics and a tool trace.

Whether anything can trade. Only Vibe-Trading, through eight brokers inside a mandate you commit on symbols, size and exposure, with a kill switch. The project calls that path experimental, and keeps its order tools off its MCP server. The other three frameworks execute nothing.

Whether you can check it. A copilot's answer is worth what its citation lets you verify. AlphaSense cites inline, down to the source document; its "no hallucinations" line is marketing. Fintool's located a document and a page, not proof of the arithmetic built on it. With a framework, what you can audit is the prompt and the tool trace, and TradingAgents' own documentation says two runs of the same ticker on the same date can disagree.

How they fail

A data tool that breaks usually breaks where you can see it: an error, an empty table, a missing field. A research agent that breaks writes a paragraph, as fluent as a good one, with nothing in it to mark the figure the model invented or the filing it never opened — Vibe-Trading on a weak model answers from training data without saying so. Read the trace or the citation before the conclusion.

Nothing here is endorsed. The ratings, convictions, price targets and "PM final call" these products emit are generated output nobody stands behind, as the projects' own disclaimers say. This catalogue describes how each one works and what it outputs, never which call to act on.

What is not here, and where it went

FinGPT has moved: it is adapters, instruction datasets and training notebooks — weights and data, not an application — and sits in backtesting frameworks under the research job, beside Kronos, FinRL and Qlib. A copilot inside a product bought for another job keeps that job's category: Fiscal.ai, Daloopa and Quartr in fundamentals research, Bigdata.com and AskNews in news and sentiment APIs. AI stock research and analysis tools sorts all of them by what the model does.

Where this category connects

The data a framework runs on is bought in market data APIs. Wiring a model to data without a research loop around it is financial MCP servers, and how to give an AI assistant market data is the setup. Financial MCP servers, and the licence behind them covers where models are weak with financial data: arithmetic over tables, point-in-time figures, and fetched documents read as instructions.