ai-research-desk

A research desk for one stock at a time. Three AI research teams study the same company (two open-source projects, TradingAgents and ai-hedge-fund, adapted here, and Edge Desk), a desk manager compares their calls for two holding periods (swing, 2 days to 8 weeks, and long term, 1 year or more), the teams debate where they disagree, and you get a short memo with a rating, the action for your position, price levels and the reasons. It is for people who pick their own stocks and want a second, third and fourth opinion that shows its work.

• It runs on your own AI plan, not an API bill. Model calls go through your Claude plan (the claude CLI) or your ChatGPT plan (the codex CLI). No per-call charges.
• Numbers come from code, not from a model. Prices, valuation and the data checks are computed. For each memo price level the model only picks a computed reference (a moving average, a recent high or low, a fair value, a Street target, optionally offset by ATR) and code works out the price; a level nothing computed supports is withheld. Figures quoted in the debate are checked against the reports.
• It refuses rather than guesses. Stale filings withhold the long-term rating for the whole desk, a missing figure is unknown rather than zero, and a failed step is shown as failed.
• Research only. It never places an order or connects to a brokerage account.
• Free and MIT licensed. No paid tier, no account, no telemetry.

Quick start · Connect your AI plan · Watch the video · Releases · Privacy

What this is not: not financial advice, not a signal service, and not a validated model. The ratings have not been shown to beat the market. Every screenshot below is synthetic demo data for fictitious companies, not anyone's real research or positions.

Privacy and your data

ai-research-desk runs on your own computer. It needs no account, sends no telemetry, and the dashboard listens on localhost only.

What stays local

• Every run: the teams' reports, the debate, the memo and the PDF, saved in memos/. Every run is kept, including several runs of one stock on the same day.
• Your settings and keys, in ~/.hedge-desk/.env.

The run files stay local, but the services below do learn which tickers you research.

External services

Your AI plan (Anthropic for Claude, OpenAI for ChatGPT): receives the prompts, which contain the ticker, prices, filings figures, headlines and the teams' reports. It never receives your keys. Whether you own the stock is kept out of every prompt that rates or debates it, so it cannot sway a rating; it is sent once per horizon to the final memo writer, which explains the action for your position (the action itself is set in code from the rating and your position). Every call runs with no tools, no web search, no file access and none of your personal CLI settings.

Alpaca (optional): receives the tickers you run, to return prices, splits and news headlines. Without Alpaca keys, prices come from Yahoo Finance, which receives the tickers.

SEC EDGAR: receives the tickers and the contact string you set in SEC_USER_AGENT, which SEC requires. It returns the companies' filings.

API keys

Keys live in ~/.hedge-desk/.env on your machine, which is never committed. Do not share or commit .env files, login folders, run memos or screenshots with personal details.

What's inside

• Three research teams: TradingAgents (analysts, a bull and bear debate, a trader and risk managers), ai-hedge-fund (investor personas over SEC filings with a research manager), and Edge Desk (a point-in-time evidence package, a deterministic business case and factor screen, then a written analysis).
• The desk manager: restates each team's call for each horizon from its own report only, sends disagreements to a debate where a manager defends or concedes (never a compromise), fact-checks what did not survive, and writes the memo. A 2 against 1 standoff is reported as gridlock, not decided for you.
• Data checks: one settled close for every team, a stale-filing rule, unknown values kept unknown.
• The dashboard: start a run, watch the teams work live, reopen any past run (a team report from earlier the same day is reused only if it came from the same plan, price session, models, fund and code), a weekly review of the latest call per symbol, and a two-page PDF summary.

Screenshots

The call. Both horizons side by side, with the action for your position, levels, who voted, how it was decided, the reasons and risks, earlier runs of the same symbol and the chart.

Weekly review. The latest call per symbol, evidence problems first, with arrows against the previous run. Here a stale filing has withheld one company's long-term rating.

The PDF summary. Two pages: the same first screen and the debates in a few lines.

Watch the video

Three AI research teams take the same stock, argue it out for a swing and a long-term call, and the desk writes up where they agree and where they do not.

Quick start

Requirements: Python 3.12 (ai-hedge-fund needs numpy 1.x, which has no wheels for newer Python on Windows), git, and a paid AI plan with its command-line tool (see Connect your AI plan).

Windows, four steps: clone the repo with git clone --recurse-submodules (or download the ZIP from Releases), then:

1. Double-click setup.bat, once. It builds one environment per part and creates ~/.hedge-desk/.env.
2. Open ~/.hedge-desk/.env (%USERPROFILE%\.hedge-desk\.env) and set SEC_USER_AGENT to your name and email. Add Alpaca keys if you have them; they are optional.
3. Log in to your AI plan once, as described in Connect your AI plan.
4. Double-click launch.bat, every time. Then open http://localhost:8790

Manual setup (Mac, Linux, or if you prefer):

git clone --recurse-submodules https://github.com/simonro/ai-research-desk
cd ai-research-desk
python3.12 install.py
# edit ~/.hedge-desk/.env: set SEC_USER_AGENT, and Alpaca keys if you have them
claude                      # log in once, if you use a Claude plan (see below for ChatGPT)
scripts/run-dashboard.sh    # then open http://localhost:8790

Every install starts empty: no runs and no demo data. Type a ticker in the dashboard, pick the teams, and start a run. A run with all three teams and a debate takes about 10 to 30 minutes.

Connect your AI plan

The desk has no API keys and no per-call bill. It sends every model call through the command-line tool of an AI subscription you already pay for, so runs count against that plan's usage. You need one of these, and you can set up both.

Claude (the default). Needs a Claude Pro or Max plan.

1. Install Claude Code, which gives you the claude command.
2. Run claude once in a terminal and sign in with your Claude account. Then close it.

That is all: DESK_PLAN=claude is the default. The desk removes any Anthropic API key from its calls, so they always go to your subscription, never to a metered key.

ChatGPT. Needs a paid ChatGPT plan that includes Codex.

1. Install the Codex CLI: npm install -g @openai/codex (needs Node.js).
2. Log the desk in once. The desk keeps its own Codex login, separate from any you already use, so your personal Codex settings never reach the analysis:

# PowerShell
$env:CODEX_HOME="$HOME\.hedge-desk\codex"; New-Item -ItemType Directory -Force $env:CODEX_HOME | Out-Null; codex login
# Mac or Linux
mkdir -p ~/.hedge-desk/codex && CODEX_HOME=~/.hedge-desk/codex codex login

3. To make ChatGPT the default, set DESK_PLAN=chatgpt in ~/.hedge-desk/.env.

Picking the plan. The dashboard's run form has a Claude / ChatGPT switch that starts on your DESK_PLAN default; changing it applies to that run only. One plan per run: a run is all Claude or all ChatGPT, and the memo records which plan and models it used.

Usage limits. A full run with all three teams and a debate makes dozens of model calls. On a smaller plan, such as Claude Pro, one or two full runs can reach your usage limit; larger plans have more room. Running one or two teams, or turning the debate off, uses less. If the limit is reached, or your login has expired, the run stops at once and saves nothing, rather than writing a half-finished memo. Log in again (claude, or the codex login command above) or wait for the limit to reset, then start the run again.

Updating to a new release

Your runs (memos/) and your settings (~/.hedge-desk/.env) are not part of a release, so an update never touches them.

If you cloned with git: git pull, git submodule update, then setup.bat (or python3.12 install.py) if dependencies changed.

If you downloaded the ZIP: unzip the new version into a new folder, copy memos/ across, and run the setup once.

Configuration

All settings live in ~/.hedge-desk/.env; .env.example documents every one.

Variable	Enables
DESK_PLAN	claude (default) or chatgpt
SEC_USER_AGENT	Required. Your name and email, which SEC asks for
ALPACA_API_KEY, ALPACA_SECRET_KEY	Alpaca prices and news (free accounts work). Without them prices come from Yahoo and there is no news
DESK_CHATGPT_STRONG, DESK_CHATGPT_FAST	The ChatGPT models for the manager and analyst roles
TRADINGAGENTS_*, HEDGE_FUND_*, EDGE_DESK_MODEL	The Claude models each team uses
FRED_API_KEY	Macro data for TradingAgents (free)
DESK_VAULT_DIR	Also save each memo as a note in a markdown vault such as Obsidian
DESK_MEMOS_DIR	Where runs are saved (default memos/)

From the command line: scripts/run-desk.sh MSFT --own --engines quant,vets,edge (.bat on Windows). --own MSFT marks it as a position you hold; a bare --own treats every ticker as new.

How far to trust it

Large language model ratings cannot be backtested honestly: the models have seen the period being tested. So the desk is forward-tested only, and nothing here has been shown to predict returns. Edge Desk prints what has been measured about its own formula, and that is a description, not a validation. Use the desk to see the arguments for and against a stock, then make your own decision.

Contributing

Bug reports and pull requests are welcome. CONTRIBUTING.md covers setup, tests and how to add a plan. Please report security problems privately, as described in SECURITY.md, not in a public issue.

More free tools from Tape to Edge

Same rules for each one: free, MIT, runs on your own machine, and the build is on the YouTube channel.

• Edge Scanner: a real-time intraday stock scanner with a browser dashboard you lay out yourself.
• Trading Journal AI: a local-first journal. Import your broker CSV, get round-trip trades, KPIs and AI diary analysis.
• Edge GEX Lab: a live gamma exposure (GEX) map for SPY, QQQ, SPX, NDX and stocks, with walls, gamma flip and regime.

Who made this

I built this to get structured, argued second opinions on the stocks I research, without paying per call on top of the AI plan I already have. Simon, Tape to Edge.

• YouTube: @tapetoedge
• X: @tapetoedge

I take no affiliate money from any tool I show, and this project has no paid tier, no account and no telemetry. If it is useful, fork it.

License

MIT. See LICENSE. TradingAgents (Apache-2.0) is included as a git submodule, and ai-hedge-fund (MIT) as a modified copy listed in engines/ai-hedge-fund/CHANGES.md.