Telegram bot wrapping the TradingAgents framework. Curate a watchlist via Telegram, tap one or more tickers, and the bot runs TradingAgentsGraph.propagate(...) for each — replying with a finviz chart, the trade decision, a short summary, and a Telegraph link to the full analysis. Per-step pipeline progress streams back into the message caption while the analysis runs.
- 🤖 Multi-Agent Analysis: Wraps TradingAgents — analyst → researcher → trader → risk-manager LLM agents collaborate on each ticker, output is a finviz chart + decision + Telegraph link.
- 📋 Watchlist Management: Curate tickers via
/add,/del,/watch. Each is yfinance-validated; class-share dot forms (BRK.B) auto-correct to dash form (BRK-B). - ⚡ Parallel Execution: Tap multiple tickers and they run in parallel, gated by
TG_BOT_MAX_CONCURRENT_ANALYSES. Overflow shows⏳ Queueduntil a slot frees. - 💾 Same-day Result Cache: Two users on the bot tapping the same ticker = one LLM run, not two.
/watch+ the daily digest hitting the same ticker = one run, not two./refresh NVDAforces a re-analysis when intraday data has shifted enough. - ⏳ Live Progress + Cancel: Per-step pipeline progress streams back into the Telegram message caption. Each in-flight analysis carries a ❌ Cancel button — cancellation is checked at every LLM-call boundary so the abort is near-instant.
- 🌅 Daily Digest:
/digestschedules a recurring run of a curated subset of your watchlist at any hour + IANA timezone — paginated multi-select picker, one summary message per day with a Telegraph link per ticker. Per-ticker market-calendar gate (exchange_calendars, 102 exchanges) skips tickers whose exchange is closed today; the whole digest short-circuits with a heads-up when every market is closed. - 📧 Email Mirror (optional): opt in via
/email set <addr>to mirror the daily digest summary to your inbox via Resend — HTML body with inline finviz charts + Telegraph links + per-ticker.mdattachments. Telegram remains the primary channel; email failures never break it. Free tier on Resend (3000/mo) covers a personal deployment. Disabled while the bot is open (emptyALLOWED_USER_IDS) — it relays through your verified Resend domain, so it only enables once the bot is locked to an allowlist; immediate/email testsends are per-user rate-limited. - 📜 Analysis History:
/historybrowses every saved analysis by ticker → date and republishes to Telegraph on demand, with← Backround-trip navigation. - 🤹 Single-tenant
.env-driven LLM config: OpenAI, DeepSeek, Anthropic, Google, xAI, Qwen, GLM, MiniMax, Ollama, OpenRouter, plus (tradingagents v0.3.0+) Mistral, Kimi/Moonshot, Groq, NVIDIA NIM, Amazon Bedrock, and a generic OpenAI-compatible endpoint. The bot reads provider + deep-think + quick-think model from.env(overlaid onto tradingagents'DEFAULT_CONFIGvia the upstreamTRADINGAGENTS_*env vars) — one shared config across every user, no in-bot picker. Optionalmax_debate_rounds(1/2/3) + per-provider reasoning-effort knobs trade cost for thoroughness; the chosen settings render under the chart caption. - 🛡 Production-grade:
ALLOWED_USER_IDSallowlist, atomic +fsyncper-user JSON storage, multi-arch Docker image (amd64+arm64) daily-rebuilt to track upstream, CodeQLsecurity-extended+ Trivy scanning, provenance + SBOM attestations.
Sample published analysis: BRK-B — 2026-05-05. Each Run from /watch produces a Telegraph page like this — chart + decision + multi-agent rationale.
/add + /watch analysis |
/digest daily summary |
Telegraph Instant View |
|---|---|---|
 |
 |
 |
bash <(curl -fsSL https://raw.githubusercontent.com/IvanWng97/TradingAgents-Telegram/main/start.sh)Prompts for your bot token, Telegram user ID, Telegraph token (auto-creates one if you skip), and an LLM provider + key. Drops a configured .env + docker-compose.yml into ./tradingagents-telegram, pulls the prebuilt image, and prints the docker-compose up -d to run next.
Prefer to do it by hand? See docs/MANUAL_INSTALL.md.
After docker-compose up -d, message your bot — /start walks you through the in-Telegram setup (/add → /watch). The active LLM config is whatever TRADINGAGENTS_LLM_PROVIDER + TRADINGAGENTS_DEEP_THINK_LLM + TRADINGAGENTS_QUICK_THINK_LLM are set to in .env — start.sh writes canonical defaults; edit them and restart to switch models. If /start replies "Not authorized. Your user ID is …" the auth gate rejected you: add that ID to ALLOWED_USER_IDS in .env, then docker-compose up -d to restart and retry.
⚠️ LeavingALLOWED_USER_IDSempty makes the bot open to anyone who finds your bot handle, and they will burn your LLM tokens. Set it.
Each analysis runs ~12 LLM calls across the agent pipeline. Per-ticker rough estimates: DeepSeek ≈ $0.01, OpenAI / Anthropic mid-tier ≈ $0.20–$0.40, frontier reasoning models ($0.50+). Tapping ✅ Done on a 10-ticker selection can easily hit single-digit dollars in minutes — pick your provider in .env accordingly. The /digest filter exists partly so you can run cheap providers across a long watchlist daily without surprise bills.
docker-compose pull && docker-compose up -dThe image is rebuilt automatically by a daily GitHub Action whenever upstream tradingagents advances; the cron skips the build when the SHA hasn't changed, so you only pull a new image when there's actually new upstream code.
| Command | Description |
|---|---|
/start, /help |
Welcome / help text |
/add NVDA AAPL TSLA |
Bulk-add tickers. Each is validated against yfinance — invalid symbols are rejected with a hint, US class-share dot forms (BRK.B) auto-correct to dash form (BRK-B). |
/add (no args) |
Bot prompts; reply with the ticker(s) you want to add |
/del NVDA AAPL |
Bulk-remove |
/del (no args) |
Inline-button picker — tap ❌ on each ticker, ✅ Done to close |
/list |
Read-only text view of your watchlist. Shows ticker count, daily digest schedule + enrolled subset (listed as → T1, T2 in the header when a proper subset is enrolled), and a monospace pre-block grid with adaptive column count (1–4 cols, narrows for long tickers like RELIANCE.NS). No buttons — use /watch to actually run an analysis. |
/watch |
Paginated select-mode keyboard (9 per page). Tap any ticker to toggle (✅ prefix), use ✓ Select all / ✗ Clear for bulk, then ✅ Done (N) runs the selected ones. Single ticker uses the cached graph; 2+ run in parallel up to TG_BOT_MAX_CONCURRENT_ANALYSES. Each in-flight analysis has its own ❌ Cancel button. |
/digest |
Schedule a daily run of a curated subset of your watchlist. Single-screen picker: 24-hour grid + 10 IANA time zones (Pacific, Eastern, UTC, GMT/BST, JST, …) + a 📋 Tickers (N/M) filter screen with a 3×3 paginated multi-select. One summary message per day, one Telegraph link per ticker. Closed-market days skip automatically — per-ticker calendar gate drops tickers whose exchange is closed (weekends + holidays via exchange_calendars); if every market is closed today, the whole digest sends a one-line "Markets closed today" heads-up instead. ▶ Run now triggers an ad-hoc fan-out; 🔕 Off disables (preserves hour, tz, and tickers for one-tap re-enable). New users start with an empty filter and must opt in. Auto-disables itself if you block the bot. |
/history (no args) |
Inline-button picker of all tickers with saved analyses |
/history NVDA |
Inline-button picker of recent analysis dates. ← Back returns to the ticker picker. |
/history NVDA 2026-04-15 |
Direct lookup — publishes that day's saved analysis to Telegraph |
/refresh NVDA |
Direct fast-path: drop today's cached result for NVDA and re-run the analysis. |
/refresh (no args) |
Same paginated multi-select picker as /watch, but the Done button is 🔄 Refresh (N) and tapping it drops today's cache for each selected ticker before running. Useful when several tickers' intraday data has shifted and you want to refresh them in one shot. |
/email [<addr>|off|test] |
Configure the daily-digest email mirror (Resend, opt-in). Bare /email shows current setting. /email foo@bar.com saves (validates locally). /email off clears. /email test sends a synthetic one-off email to verify wiring. Requires the operator to set RESEND_API_KEY + RESEND_FROM in .env and a non-empty ALLOWED_USER_IDS (disabled in open mode; off always works). /email test is per-user rate-limited. |
/status |
Operational snapshot: uptime, # analyses since boot, graph-pool size, active LLM config (provider + deep + quick + rounds + effort + temperature when set) loaded from .env |
The Telegram client also exposes a Menu button next to the input field with the same commands (populated via set_my_commands), and /-autocomplete works after typing the first letter or two.
All env vars — required secrets and tuning knobs — live in docs/CONFIGURATION.md. The four required secrets are pre-templated in .env.example.
Common startup / runtime issues and fixes: docs/TROUBLESHOOTING.md.
docs/MANUAL_INSTALL.md— manual Docker setup if you'd rather not runstart.sh.docs/CONFIGURATION.md— env vars (required + tuning) and supported LLM providers.docs/DEVELOPMENT.md— local dev setup, pre-commit, smoke tests.docs/TODO.md— roadmap items not yet started.CLAUDE.md— architecture notes, key contracts, and current limitations.
This bot wraps the TradingAgents research framework — output is generated by LLMs and is not investment advice. Decision quality depends on the provider + deep / quick models set in .env, the yfinance data window the agents see, and the day's market noise; agents can hallucinate, contradict themselves, or miss context the prompts didn't surface. The same-day cache reuses a morning run for later taps on the same ticker — if intraday data has shifted, use /refresh. Use as a research aid only, do your own due diligence before acting on any signal, and don't run /digest on tickers you'd let it trade unsupervised. The author and contributors accept no liability for decisions made using this software.