FinRobot

What it is

FinRobot is an open-source multi-agent stack for equity research from the AI4Finance Foundation, the academic group behind FinGPT and FinRL. Give it a ticker and your own OpenAI and Financial Modeling Prep keys and it pulls the statements, builds three-year forecasts and a DCF, runs eight LLM agents over the result, and renders an analyst-style HTML or PDF report — from a local FastAPI UI on 127.0.0.1:8001 or from two command-line scripts.

Which half is open matters. The repo holds V0, the 2024 AutoGen agent library, and V1, finrobot_equity, the report generator. V2 is a hosted product at finrobot.ai behind a sign-in and its source has not been released; FinRobot Desktop v0.1.0 (7 July 2026, Apple Silicon only, not notarised, so first launch needs xattr -cr) is a binary on that same closed stack.

Apache-2.0, roughly 8,000 stars, not archived. Last release v1.0.0 on 20 March 2026; the last code change of any kind landed 11 May 2026 — a Docker and Cloud SQL deploy setup — and all eleven commits since then touch only the README, the most recent on 11 September 2026.

Pricing

Nothing to buy that we could find. The repo is free, and finrobot.ai publishes no price at all — /pricing, /plans, /billing and /subscribe all 404, the site root and /v2 both redirect to a sign-in, and sign-up asks only for a full name, an email and a password (checked 13 September 2026). The bill is elsewhere: an OpenAI key for every agent call, an FMP key for the financials, both yours.

Data & coverage

It ships no data. V1 reads statements, ratios, quotes, the analyst rating and the price target from FMP, and historical volume from yfinance; V0 also wires up Finnhub, SEC EDGAR via sec-api and Reddit via FinNLP. The NVDA, MSFT, COP, TSLA and META samples in the repo were generated by that code on 21–22 March 2026 and show what comes out: a cover page with a rating and a price target, then thesis, financials, valuation, peers and risks.

Integrations

V0 wraps backtrader in one function, so an agent can run and plot a backtest over yfinance bars — SMA crossover by default, or any strategy, sizer and indicator class you name, with Sharpe, drawdown and trade analyzers attached. There is no MCP server in the repo and no broker connection anywhere in it.

Limitations

• The report opens on a Buy/Hold rating and a target price. In V1 those come from FMP's upgrades-downgrades and price-target endpoints, not from the model — FinRobot reprints someone else's call under its own masthead — while the LLM writes the thesis around them. The vendor's architecture page has step 7 of the V2 pipeline combine everything "into an investment thesis and a rating". Either way, nobody signs it.
• pip install finrobot installs 0.1.5 from June 2024 — the V0 library, not the report generator.
• setup.py and the PyPI classifiers say MIT; the LICENSE file and NOTICE say Apache-2.0. The published wheel carries the contradiction inside itself — its metadata reads License: MIT beside a dist-info/LICENSE that is the Apache 2.0 text.
• V1 is built on the OpenAI SDK and Agents SDK and pinned to openai<2 and numpy<2, though openai_base_url accepts any OpenAI-compatible endpoint.
• US-only in practice: every figure is formatted in dollars, "How to retrieve non-US listed stock?" has been open since June 2024, and a July 2026 request to document non-US ticker support is open too.
• Two path-traversal code-injection reports against 1.0.0, in CodingUtils.create_file_with_code and modify_code, were filed on 28 May 2026 — both functions join a model-supplied filename onto a fixed prefix and open it. Both are open with no maintainer reply, and an outside pull request that fixes them by resolving paths against a workspace boundary has sat unmerged since 10 June 2026. Sixty issues are open in total.
• Not a trading system: no broker, no order routing, no portfolio tracking.

Alternatives

Vibe-Trading and OpenBB cover much more ground and both ship an MCP server, which FinRobot does not. For filings without an agent wrapped around them, SEC EDGAR MCP does that one job. If the valuation is what you want rather than the prose, a spreadsheet and a data API are cheaper and do not invent a target price.