Repository navigation
Providers
Short version: one LLM key makes the agents smart, and four free provider keys unlock the full experience: sentiment, fundamentals, forecasts, and paper trading. They are not nice-to-haves, grab all of them; the whole set takes about ten minutes.
No key at all? The app still runs in mock mode: the whole workflow works on live market data with rule-based agents, clearly labeled so you know what you're looking at. Great for a first look.
Click a link, create the account, copy the API key it hands you, paste it in. That's it.
| Provider | Click here to sign up and grab your key | Why pick it |
|---|---|---|
| Z.AI (GLM) | opens the Z.AI API key page | Cheap and fast. GLM-5.3-Flash is a great everyday default |
| DeepSeek | opens the DeepSeek API key page | Solid reasoning at low prices |
| OpenRouter | opens the OpenRouter key page | One key, hundreds of models to swap between |
| Alibaba Qwen | opens Model Studio, where the DashScope key lives | Fast models with big context windows |
| Fireworks AI | opens the Fireworks API key page | Hosted open models (Llama 4, DeepSeek, Qwen, Kimi) at pay-per-token rates; new accounts get $1 in free credits. Model IDs use the full form accounts/fireworks/models/<model>
|
| OpenAI | opens the OpenAI API key page | The classic; usually the priciest option |
Every one of those pages walks you through creating an account first if you don't have one; the key is a string like sk-... on the same screen.
Two jobs, two tastes. Picks below track the Artificial Analysis Intelligence Index as of October 2026, so "really good" means measured, not vibes.
For fast agent work (the researchers and the debate), we recommend one of:
| Model | Model id | Why |
|---|---|---|
| Claude Haiku 5.5 |
anthropic/claude-haiku-5.5 (OpenRouter) |
AA #20 (43.4); 382 tok/s at $0.10/$0.50 per 1M tokens, the cheapest credible fast tier |
| GPT-6 Luna | gpt-6-luna |
Fast OpenAI pick for agent loops |
| GLM-5.3-Flash | glm-5.3-flash |
Great everyday default, very cheap |
| DeepSeek 4.1 Flash | deepseek-4.1-flash |
AA #23 (39.5); 320 tok/s, a latency record |
For reasoning and coming up with a good report (the manager's final call), we recommend one of:
| Model | Model id | Why |
|---|---|---|
| Claude Opus 5.5 |
anthropic/claude-opus-5.5 (OpenRouter) |
AA #1 (57.6), the top score measured; $4/$20 per 1M tokens |
| Claude Sonnet 5.5 |
anthropic/claude-sonnet-5.5 (OpenRouter) |
AA #2 (56.0); near-Opus at half the price |
| GPT-6.1 Sol | gpt-6.1-sol |
OpenAI's reasoner (GPT-6 Astra sits at #7, 52.7, but costs $10/$50) |
| Gemini 4 Argon | Gemini API or OpenRouter | AA #8 (52.6) at $2/$10 |
| MiMo-V2.6-Pro | open weights; hosts listed on its AA model page | AA #15 (46.3); beats pricier closed APIs at ~$0.44/$0.87 |
| GLM-5.3 | glm-5.3 |
Strong reasoning at a fraction of the frontier price |
Put the fast one in LLM_MODEL and the strong one in LLM_MODEL_MANAGER (Settings page or .env), so the crew runs cheap and the judgment runs deep. Model ids shift over time: copy the exact id from your provider's console. A typical analysis costs a few cents.
These are not nice-to-haves: Finnhub, Olostep, and Nixtla feed the analysts, and Alpaca is what turns decisions into simulated trades. All four are free to start.
| Key | Click here to sign up | What it unlocks |
|---|---|---|
| Finnhub | straight to the free registration | Company fundamentals and news (falls back to yfinance without it) |
| Olostep | homepage, then Sign up; free credits to start | Reddit and StockTwits sentiment, plus a news-search fallback |
| Nixtla TimeGPT | straight to the dashboard login/signup, free credits included | A 5-day AI price forecast (falls back to a local trend model) |
| TSFM.ai | homepage, then Sign up and mint a key under Account, API Keys | Hosted Chronos-2 time-series forecast; used when TimeGPT is unavailable. Free credits on signup |
| Alpaca paper | straight to the free account signup | Turns decisions into simulated orders and makes the Trades page live. Your API keys are in the paper trading section of the dashboard |
Two ways, both fine:
- Run
uv run setup: the wizard asks for everything, hides your key while you type, and can test it with one real request. - Launch the app and open Settings: keys saved there are stored in your OS keychain (Windows Credential Manager and friends), not in a file.
Missing one key while you set up? No drama: the app falls back gracefully and tells you which provider answered what, and thin or missing data reads as "not a signal" instead of being invented. But for the real experience, get all of them.