AI Hedge Fund

What it is

aihf is a terminal app and importable Python package that assembles a fake fund out of alpha models and replays it over history. Five of those models are LLM investor personas — Buffett, Munger, Graham, Lynch, Druckenmiller — and one, post-earnings-announcement drift, is arithmetic. All return the same object: a conviction in [-1, +1] plus a written thesis. Deterministic code, not the language model, blends those views into target weights, clamps them against per-position and gross-exposure limits, and sends orders to a simulated broker.

A persona is a system prompt and nothing more — buffett.py is 54 lines, most of it prose. The engineering is elsewhere: the fundamentals snapshot an agent sees is filtered on filing date rather than report period, and every prompt and response is cached by content hash, so an unchanged snapshot never pays for a second call.

Version 2 deleted v1's React frontend and LangGraph pipeline outright. Last release v2.2.0 on 7 August 2026, last commit to main on 3 September 2026, MIT, not archived, 63.4k stars.

Pricing

Free software, paid inputs. Data comes from Financial Datasets and only from there; that API has no free tier as of today, starting at $20 one-time for 1,000 requests. Then one LLM key. Because the prompt cache keys on the filing snapshot, a weekly backtest re-reasons only when a new filing lands — the snapshot is trailing-twelve-month rows, so roughly four calls per persona-ticker-year rather than one per rebalance date.

Data & coverage

US equity fundamentals and daily bars, filed-date-aware, cached to ~/.hedge-fund/cache. No intraday, no options, no futures, no crypto. The mandate's benchmark, SPY by default, sets the trading-day grid.

Integrations

Installed as pipx install aihf, and the name matters. The obvious guess, pip install ai-hedge-fund, is a different author's package — a 0.1.1 set of Claude Code slash commands uploaded in April 2026. hedge-fund is this author's own, abandoned at 2.1.0 three days after he published it. Only aihf — 2.2.0, uploaded 7 August 2026 — is this project.

It wants two keys, asked for on first use and written to ~/.hedge-fund/.env, with shell variables overriding the file: one for Financial Datasets, one for any of the nine models across six providers — Anthropic, OpenAI, Google, DeepSeek, xAI, Kimi — listed in hedge_fund/llm/api_models.json.

The data source and the broker are both Python Protocols, so a replacement needs no inheritance; the data protocol's docstring sketches a yfinance client as its example. Neither has a second implementation in the tree. There is no MCP server, no Dockerfile and no web UI. A non-interactive run prints the whole cycle record as JSON on stdout and a human summary on stderr, and a fund is a YAML mandate under ~/.hedge-fund/mandates/.

Limitations

• Nothing executes. SimBroker is the only broker; paper and live are roadmap items.
• The backtest models no trading costs: full fills at the reference close, no commission, slippage or spread, and no margin — cash is allowed to go negative and stay visible.
• Each persona returns bullish, bearish or neutral with a 0–100 confidence on a named security. That is a rating on a stock, whatever the disclaimer says, and the project's own answer is that it is for education and research only.
• Personas see fundamentals alone — no prices, no news, no transcripts.
• The always-on fund the vision describes is not built: no scheduler, and NAV resets to the mandate's capital each run because the ledger's read half is unfinished.
• 53 open issues and 109 open pull requests on 14 September 2026, a good number of both written against v1 code that no longer exists.

Alternatives

TradingAgents and FinRobot run multi-agent debates over more data types; Vibe-Trading wires an agent to real brokers. For a backtester with no LLM in it at all, see the backtesting frameworks category, and for the rest of this one, MCP servers and AI agents.