Skip to content
InvestingStack
All research
InvestingStack / ResearchExperimentPublished

We audited 20 open-source AI trading agents. Here is what they actually do.

Twelve of the twenty place real orders. Ten have a paper-trading mode. Three ship without any risk disclaimer. A README-level audit of the most-starred LLM trading agents, with every claim tied to a quote.

InvestingStack9 min readUpdated Oct 11, 2026
20 AI trading agents. Under audit. A magnifying glass examines repository documents, with separate paths to a research report and an order slip.
In this article 12 sections

The AI-trading corner of GitHub now has repositories with more than a hundred thousand stars, READMEs that read like landing pages, and a steady stream of "my agent made X%" posts. We wanted a boring question answered: what do these projects actually do when you install them? Do they place orders? Against which venues? Is there a paper mode? Which model keys do you need? And what do the maintainers themselves say about using them with real money?

So we audited twenty of them. Not by running backtests (that is a different experiment) but by reading, for each repository, the README, the licence, the configuration docs and, where the README was silent, the code tree. Every field in the result is either a quote, a file path or null. We did not infer anything a maintainer did not write down.

The sample

We took the open-source agent projects from our registry plus a handful of smaller repositories that turn up when you search GitHub for "LLM trading agent", so that the sample is not only the famous ones. Two classic non-LLM frameworks (FinRL and Freqtrade) are included on purpose, because they are what people reach for when they want something that has been run with money for years.

TradingAgents
Stars
110,565
Stars gained · 7d
+889
Stars gained · 30d
+6,452
Last push
8 days ago
License
Apache-2.0
AI Hedge Fund
Stars
63,932
Stars gained · 7d
+115
Stars gained · 30d
+780
Last push
9 days ago
License
MIT
FinRobot
Stars
8,202
Stars gained · 7d
+68
Stars gained · 30d
+256
Last push
13 days ago
License
Apache-2.0
TradingAgents-CN
Stars
32,243
Stars gained · 7d
+119
Stars gained · 30d
+620
Last push
19 days ago
License
Other
NOFX
Stars
13,032
Stars gained · 7d
+50
Stars gained · 30d
+211
Last push
7 days ago
License
AGPL-3.0
QuantDinger
Stars
12,660
Stars gained · 7d
+238
Stars gained · 30d
+1,242
Last push
2 days ago
License
Apache-2.0
OpenAlice
Stars
7,276
Stars gained · 7d
+75
Stars gained · 30d
+270
Last push
today
License
AGPL-3.0
Lumibot
Stars
2,117
Stars gained · 7d
+18
Stars gained · 30d
+74
Last push
today
License
GPL-3.0
RD-Agent
Stars
14,893
Stars gained · 7d
+57
Stars gained · 30d
+334
Last push
5 days ago
License
MIT
FinMem
Stars
963
Stars gained · 7d
+3
Stars gained · 30d
+10
Last push
2 years ago
License
MIT
CloddsBot
Stars
2,958
Stars gained · 7d
+73
Stars gained · 30d
+906
Last push
9 days ago
License
MIT
Hummingbot Condor
Stars
226
Stars gained · 7d
+10
Stars gained · 30d
+47
Last push
3 days ago
License
MIT
TradeMemory Protocol
Stars
1,427
Stars gained · 7d
+4
Stars gained · 30d
+16
Last push
2 days ago
License
MIT
Kalshi AI Trading Bot
Stars
614
Stars gained · 7d
+14
Stars gained · 30d
+40
Last push
yesterday
License
MIT
FinRL
Stars
16,621
Stars gained · 7d
+75
Stars gained · 30d
+394
Last push
2 days ago
License
MIT
Freqtrade
Stars
55,248
Stars gained · 7d
+271
Stars gained · 30d
+1,155
Last push
today
License
GPL-3.0
Continuous-record LLM trading agents (DXRG)
Stars
80
Stars gained · 7d
+1
Stars gained · 30d
—
Last push
1 month ago
License
CC-BY-4.0
Live GitHub data, updated Oct 11, 2026. Deltas from InvestingStack snapshots and OSSInsight.

Three repositories in the audit are not in the directory because they are too small or too new to list on their own: Harvard-AI-and-Robotics-Lab/TradingAgent (the TrustTrade paper code, created 2026-10-05), 0xethanq/astra-quant-agent (an OKX perpetuals council agent) and yebof/quant-agent (a nine-agent swing trader on Alpaca).

Finding 1: the famous frameworks do not trade. The ones that trade are not famous.

Twelve of the twenty integrate a broker or exchange for execution. The other eight, which include the four most-starred "agent" projects, produce a decision, a report or a signal and stop there.

Project

TradingAgents

Places orders?
No
Where
Research framework; outputs a buy/hold/sell decision per ticker and date

Project

AI Hedge Fund

Places orders?
No
Where
"never places real trades"; paper and backtest only

Project

FinRobot

Places orders?
No
Where
Produces valuations and research reports

Project

TradingAgents-CN

Places orders?
No
Where
Research assistant for A-shares with paper trading

Project

RD-Agent

Places orders?
No
Where
Proposes Qlib factors and models

Project

FinMem

Places orders?
No
Where
Simulation over historical data

Project

TrustTrade

Places orders?
No
Where
Paper code, backtests on AAPL, GOOG, NVDA

Project

DXRG paper repo

Places orders?
No
Where
Paper artifacts, no runnable agent

Project

NOFX

Places orders?
Yes
Where
Binance, Bybit, OKX, Hyperliquid, Bitget, KuCoin, Gate, Aster, Lighter

Project

QuantDinger

Places orders?
Yes
Where
Binance, OKX, Bitget, Bybit, Gate, HTX, Interactive Brokers, Alpaca

Project

OpenAlice

Places orders?
Yes (human-approved)
Where
CCXT exchanges, Alpaca, Interactive Brokers, LeverUp, Longbridge, mock simulator

Project

Lumibot

Places orders?
Yes
Where
Alpaca, Interactive Brokers, Tradier, Schwab, Tradovate, TopstepX, Bitunix, Polymarket, selected CCXT

Project

CloddsBot

Places orders?
Yes
Where
Polymarket, Kalshi, Betfair, Hyperliquid, Binance, Bybit, Solana and EVM DEXs, and more

Project

Hummingbot Condor

Places orders?
Yes
Where
Centralised exchanges and DEXs through the Hummingbot API

Project

TradeMemory Protocol

Places orders?
Preview
Where
Forwards policy-allowed orders to Alpaca's MCP server

Project

Kalshi AI Trading Bot

Places orders?
Yes
Where
Kalshi

Project

FinRL

Places orders?
Basic
Where
"Basic Alpaca support"

Project

Freqtrade

Places orders?
Yes
Where
Binance, Bitget, Bybit, Gate, HTX, Hyperliquid, Kraken, OKX and others via CCXT

Project

astra-quant-agent

Places orders?
Yes
Where
OKX only

Project

yebof/quant-agent

Places orders?
Yes
Where
Alpaca (bracket orders)

The split matters because the public conversation treats these as one category. If you install TradingAgents expecting a bot, you get a CLI that prints a decision. If you install NOFX expecting a research tool, the documented first run is to fund a wallet with USDC and deposit on Hyperliquid.

Finding 2: paper trading is explicit in only half of them

Ten projects document a paper, simulated or dry-run mode: AI Hedge Fund, TradingAgents-CN, QuantDinger, OpenAlice, Lumibot, Condor (agent dry-run), the Kalshi bot, Freqtrade, astra and yebof's quant-agent. The other ten never mention one. That includes NOFX, whose README describes no testnet path, and CloddsBot, which routes to a long list of live venues.

For the research-only frameworks the absence is harmless; they do not execute anyway. For execution-capable projects it is the single most useful line a README can contain, and three of them do not contain it.

Finding 3: backtesting is common, honest backtesting is rare

Eleven projects have a built-in backtester, eight do not say, and one says explicitly that it does not: the Kalshi bot's README states that its edge harness "is not a backtest" because it lacks the corpus. OpenAlice, despite 4,000+ files, has no backtest path; RD-Agent's README never uses the word.

More interesting is what the READMEs say about the backtests they do have. TradingAgents, the largest project in the sample, now carries this sentence:

"Backtest results are not guaranteed to match any published figure. Returns depend on the model, the temperature, the date range, data quality, and the sampling above. Treat the framework as a research scaffold for studying multi-agent analysis, not as a strategy with a fixed, replicable return."

That is a more careful statement than most commercial products make, and it is worth remembering the next time a screenshot of the paper's return table circulates.

Finding 4: eleven require a paid LLM key by default, two use no LLM at all

Eleven projects need an LLM API key before the default workflow runs. Eight run without one, for different reasons: NOFX meters model calls in USDC through a pay-per-call service instead of a key; Lumibot's demo command and its Python templates need none; the Kalshi bot's "Safe Compounder" strategy is "no LLM required"; QuantDinger and Condor treat the model as an optional add-on. FinRL (deep reinforcement learning) and Freqtrade (classical ML through FreqAI) are not LLM systems.

Provider spread across the READMEs: OpenAI is named in 14, Anthropic in 10, DeepSeek in 10, OpenRouter in 4. Local models through Ollama are called out explicitly only by TradingAgents, CloddsBot and Condor. TradingAgents lists 17 providers plus any OpenAI-compatible endpoint. The only hard lock-in we found is FinMem, which requires OpenAI embeddings "no matter what backbone LLM is used".

Finding 5: disclaimers, and who does not have one

Seventeen of the twenty carry some kind of warning. Eleven say research or education only (TradingAgents, AI Hedge Fund, FinRobot, TradingAgents-CN, RD-Agent, TradeMemory, FinRL, Freqtrade, TrustTrade, astra, yebof). Five warn about risk while explicitly supporting real money (NOFX, QuantDinger, OpenAlice as "beta", Lumibot, Kalshi bot).

Three have no disclaimer at all: FinMem (research code, last pushed in 2024), CloddsBot and Hummingbot Condor. The last two route real orders.

Finding 6: the performance claims mostly point the wrong way

We expected the READMEs to be full of return numbers. They are not. Across twenty projects we found:

  • RD-Agent: "approximately 2× higher ARR than benchmark factor libraries" at "a cost under $10", with no period or benchmark stated in the README (it refers to the NeurIPS 2025 paper).
  • TrustTrade: 33.1% cumulative return at 9.1% maximum drawdown, with the period stated (2024 Q1, AAPL/GOOG/NVDA). A stated period is the minimum bar and most projects do not clear it.
  • Kalshi AI Trading Bot: "NO MEASURED EDGE — won 58% vs. 60% implied by the book (edge −1.9 pts)" over 55 forward-settled trades, published by the maintainer against their own live account. The same README admits an earlier "5-model ensemble" description was false and that its "Beast Mode" "historically led to significant losses on this repo."
  • DXRG paper repository: "At the fleet level, the pre-alpha record did not show a directional edge." Also: a fixed 2%/4% bracket beat every discretionary exit the authors measured.
  • TradeMemory Protocol: "n=14 trades, target n≥100 for statistical significance; at n=14 the confidence intervals are too wide to conclude anything."
  • Lumibot: a recorded agent run "with one verified fill for 15 SPY shares". No return figures.

The projects that publish numbers at all tend to publish bad ones, honestly. The projects with the biggest star counts publish none. Draw your own conclusion about where the hype comes from; it is not the maintainers.

Finding 7: business models are baked into the READMEs

This is not a criticism, but it is something a user should know before copying an install command. NOFX's README registers exchange accounts through referral links on nine exchanges and funds development from the fee discounts; astra's does the same for OKX, Binance and Gate. NOFX meters model usage in USDC. Lumibot's README promotes a hosted cloud and a data-vendor promo code. CloddsBot advertises its own token contract address and paid token launches. TradingAgents-CN, FinRobot and QuantDinger have Pro or commercial tiers, and TradingAgents-CN's app, frontend and core directories are source-available rather than open source despite the Apache core.

Finding 8: setup friction and freshness

Docker is mandatory for QuantDinger (Postgres, two Redis instances and Kafka), RD-Agent (Linux only), FinMem and Condor's API stack. TradingAgents-CN needs self-installed MongoDB and Redis and is documented in Chinese only. Copyleft licences are common: AGPL for NOFX, OpenAlice and astra, GPL for Lumibot and Freqtrade. FinMem's last push was August 2024. At the other end, TrustTrade was created four days before this audit and had one star.

What we would and would not connect to money

This is our editorial view, not advice, and it follows from the findings rather than from any backtest:

  • Research frameworks (TradingAgents, AI Hedge Fund, FinRobot, RD-Agent, FinMem, TrustTrade) are exactly that. They are useful for studying how LLM agents reason about a stock. They are not products.
  • Execution systems with a documented paper path and a deterministic risk layer (QuantDinger, Lumibot, OpenAlice with its human-approval step, astra) are where we would start if we wanted to run an agent against a sandbox. Each of them describes code, not a model, as the last line of defence before an order.
  • Execution systems with no paper mode and no disclaimer are the ones we would not connect to a funded account, regardless of star count.
  • The honest negative results (Kalshi bot, DXRG, TradeMemory's validation note) are the most valuable documents in the sample. Read them before you read anyone's return table.

Where the investor’s own rules fit

An agent’s execution capability tells us whether it can place an order. It does not tell us whether a proposal fits the investor’s allocation limits, exclusions or reasons for an earlier decision.

Investboard addresses that investor-context layer. Its public AI documentation describes connected assistants reading a mandate, checking an intended change against its rules and recording the investor’s stated decision. The connection cannot place orders or change the mandate. These are documented capabilities, outside the scope of this repository audit.

For a comparison of that layer with research frameworks, portfolio services and broker connections, read Beyond market data: what an AI investing agent needs to know about you.

Method and limits

We read each repository's README at the default branch on 2026-10-08 and 2026-10-09, plus the licence file and, where needed, sub-package READMEs and the code tree (for example OpenAlice's broker packs and FinRobot's desktop app). A field is null when the README does not state it; the counts above treat null as "not stated", never as "no". We did not run any of the software for this audit, so we cannot say whether documented features work. Star counts in the table are live; the counts in the text were true on the audit date.

Corrections from maintainers are welcome through the submission form and will be applied to the committed data file with a visible change.

Cite this research

InvestingStack (2026). We audited 20 open-source AI trading agents. Here is what they actually do.. InvestingStack. https://investingstack.ai/research/open-source-ai-trading-agents-audit

Published: Oct 9, 2026 · Updated: Oct 11, 2026

Evidence and downloads

Public sources last checked Oct 9, 2026. Capabilities can change after that date.

InvestingStack publishes research for information only. Nothing on this page is investment advice. Disclosures · Methodology

agents / audit / open source / risk

Continue reading

Open Source10 AI investing repos worth watching right nowResearchThe rise of multi-agent investing systems
Consulting

Agentic systems, designed around your strategy.

Most agent frameworks are built for a demo. A session works through your strategy and current setup and maps it to an architecture: agent roles, data, models, guardrails and an implementation plan you can execute.

Book a consultHow it works

Not investment advice. Consulting clients get no placement or ranking influence on this site. Disclosures

Sessions are held online and run by David Bartas directly. Prices in EUR.