PRISM-INSIGHT
An open-source AI stock analysis and trading system where 13+ specialized agents detect surge stocks, produce analyst-grade reports, and execute trades automatically.
Evidence shows secrets managed via env vars/secrets.yaml with a test asserting no inline secrets (sensitive data 2); dependencies pinned with annotated rationale and a license-compliance CI job (dependency security 2); THIRD_PARTY_NOTICES.md plus compliance check support attribution (2). Deductions: the system executes real brokerage orders, Telegram broadcasts, and external API fetches by default, and no file shown demonstrates a user-confirmation gate before orders or a stated permission boundary; rollback/recovery is supported only by test filenames (runner_recovery, broker_recovery) with no code shown, so least_privilege, user_confirmation, external_effects, and rollback all score 1; data flows to OpenAI/Perplexity/Firecrawl are not documented in the provided files.
Docs and CI are mutually consistent, tests span Python 3.10–3.12, and offline safety regressions are well organized (self_consistency 2). Deductions: the core mcp-agent dependency is the author's own fork pinned to a commit with an unresolved upstream PR TODO, leaving upstream availability suspended (1); failure handling appears only as scattered notes (JSON repair, KRX circuit breaker) with no systematic error-message discipline (1).
KR/US dual markets with quickstart, demo.py, Docker, and multi-language reports give solid audience coverage (2); CI matrix and Docker support environment fit (2). Deductions: capability boundaries are limited to the investment disclaimer and the 'figures are diagnostic' note; no file shows when trading agents refuse to act (1); surge-trigger precision depends on TRIGGER_BATCH_ALGORITHMS.md, which is not among the provided evidence (1).
Clear documentation index, five-language README, thorough install notes, ample examples/videos, honest disclaimers and a 'not a realizable backtest' caveat, full AGPL-3.0 text with documented dual licensing, CLA and maintainer contact — these support scores of 2. Deductions: a legacy/native dual-format config transition burdens naming stability (1); version evolution (v1.16.7–v2.18) is referenced but no CHANGELOG file is present (1).
Diverse outputs (PDF, Telegram, dashboard, mobile app) support usability (2); the 13-agent pipeline with a self-improving trading journal offers differentiated value (2). Deductions: ~$310/month operating cost is self-reported but user-side API cost versus benefit is unquantified, and headline returns (+244.63%) cannot be verified inside the repository (cost_benefit 1).
The README explicitly separates diagnostic statistics from realizable returns and disclaims historical backfill — good fact/inference separation (2). Deductions: headline performance claims point only to an external dashboard with no raw data or reproduction scripts in the repo (traceability 1); cross-source corroboration exists only at that external site and cannot be settled from these files (1).
- The system supports real brokerage auto-trading; this static review found no code evidence of a mandatory pre-order user-confirmation gate. Audit the execution path yourself and start with small amounts.
- Default modes broadcast to Telegram channels and send market data to OpenAI, Perplexity, and Firecrawl; assess your own exposure for leaking sensitive strategy and position information.
- The core mcp-agent dependency is the author's private fork (pinned commit, upstream PR unmerged), creating long-term maintenance and security-update uncertainty.
- Cumulative return and win-rate figures in the README are operator-reported and point to an external dashboard; they cannot be verified in this repository and are officially labeled non-time-weighted — do not trade on them.
- Commercial SaaS use requires a separate paid license; verify your AGPL-3.0 obligations, including source disclosure for network services.
What does this agent do, and when should you use it?
PRISM-INSIGHT is an open-source (AGPL-3.0), AI-powered stock analysis and trading system by dragon1086 on GitHub, covering the Korean (KOSPI/KOSDAQ) and US (NYSE/NASDAQ) markets. Its core is a multi-agent workflow: 13+ specialized agents organized into Macro, Analysis, Strategy, Communication, Trading, and Consultation teams collaborate on surge detection, technical/financial/industry/news/market analysis, strategy synthesis, and buy/sell decisions. Data flows through MCP servers (kospi_kosdaq for KRX, yahoo-finance-mcp and sec-edgar-mcp for US markets), with OpenAI GPT-5 as the default model and a Codex OAuth Proxy that lets ChatGPT Plus/Pro subscribers run the pipeline without an API key. Outputs include multilingual PDF reports, Telegram broadcasts, trading simulations, and real execution via the Korea Investment & Securities API, backed by a public live dashboard and iOS/Android apps. It fits quantitatively-minded users who can self-host, absorb API costs, and verify trading risk themselves.
The orchestrator (stock_analysis_orchestrator.py; prism-us/us_stock_analysis_orchestrator.py for US markets) runs in morning/afternoon modes: it detects stocks with unusual volume/price movement to build a watchlist, runs six Analysis agents (technical, financial, industry, news, market) in parallel, synthesizes strategy, and has Trading agents make buy/sell decisions recorded in a trading journal whose trigger win-rate statistics feed back into future buy decisions (a self-improving loop). Data comes via MCP servers: KRX data (kospi_kosdaq), web crawling (firecrawl), web search (perplexity), simulation DB (sqlite), US OHLCV/financials (yahoo-finance-mcp), and SEC filings/insider trading (sec-edgar-mcp). Outputs are analyst-grade PDF reports rendered with Playwright/Chromium, multilingual Telegram broadcasts, and real order execution through the Korea Investment & Securities API. An optional ADANOS_API_KEY enriches US stock news with structured social sentiment. A separate Stance leaderboard records strategy decisions in real time from registration day, forbids historical backfill, and ranks by return, worst drawdown, average invested exposure, and record rate.
- A Korean retail investor who wants daily AI analysis of KOSPI/KOSDAQ surge stocks delivered automatically to Telegram instead of watching the market manually.
- A US-stock researcher with only an OpenAI API key using the 60-second quickstart: ./quickstart.sh generates a PDF report for AAPL, MSFT, NVDA, or TSLA.
- A ChatGPT Plus/Pro subscriber with no API budget running the full pipeline via the Codex OAuth Proxy (PRISM_OPENAI_AUTH_MODE=chatgpt_oauth).
- A systematic-trading developer registering their own strategy on the Stance leaderboard by pasting one instruction into Codex CLI, Cursor, or Claude Code, entering official ranking after 63 trading days and 20 closed trades.
- A user who wants real automated execution via the Korea Investment & Securities API after configuring mcp_agent.secrets.yaml and KRX credentials.
- A multilingual content operator using the built-in translation agent to publish EN/JA/ZH/ES reports through dedicated Telegram channels.
What are this agent's strengths and limitations?
- Covers the full end-to-end flow — surge detection, multi-agent analysis, trading simulation, and real order execution — not just report generation.
- Dual-market support (KRX and NYSE/NASDAQ) with US data from yahoo-finance-mcp and sec-edgar-mcp including SEC filings and insider trading.
- A Codex OAuth path lets ChatGPT Plus/Pro subscribers run the full analysis with zero API bills.
- The trading-journal feedback loop (win-rate tracking informs future buy decisions) and a public live dashboard provide verifiable operating records.
- Rich ecosystem: iOS/Android apps, Telegram channels in five languages, and a no-backfill Stance strategy leaderboard.
- Operation depends on explicitly listed paid external services: ~$235/month for OpenAI API and ~$310/month total including Anthropic, Firecrawl, Perplexity, and server infrastructure.
- Korean-market live execution is tied to the Korea Investment & Securities API and KRX/Kakao credentials, requiring adaptation for non-Korean users at that step.
- Analysis depth is anchored to OpenAI GPT-5; Claude is only an optional compatibility workflow, and migration to other models is not documented.
- Performance figures in the README are diagnostic: the Korean-market +244.63% cumulative return is a sum of trade-level rates, not a time-weighted portfolio return or a realizable backtest.
- Full setup involves Python, Node.js MCP builds, Playwright, multiple YAML config files, and KRX credentials — a non-trivial deployment cost.
How do you install or deploy this agent?
Quickstart (US stocks, only an OpenAI API key needed):
git clone https://github.com/dragon1086/prism-insight.git
cd prism-insight
./quickstart.sh YOUR_OPENAI_API_KEYOr with Docker:
export OPENAI_API_KEY=sk-your-key-here
docker compose -f docker-compose.quickstart.yml up --build -d
docker exec -it prism-quickstart python3 demo.py NVDAFull install (Python):
pip install -r requirements.txt
python3 -m playwright install chromiumcd perplexity-ask && npm install && npm run build && cd ..
cp mcp_agent.config.yaml.example mcp_agent.config.yaml
cp mcp_agent.secrets.yaml.example mcp_agent.secrets.yaml
# Add your OpenAI API key to secrets; add KRX credentials (Kakao account) to configOptional: a Perplexity API key for news analysis and ADANOS_API_KEY for US news sentiment. Full guide: docs/SETUP.md.
How do you use this agent?
Run analysis:
python stock_analysis_orchestrator.py --mode morning --no-telegramUS markets:
python prism-us/us_stock_analysis_orchestrator.py --mode morning --no-telegram
python prism-us/us_stock_analysis_orchestrator.py --mode morning --language enSingle-stock demo: python3 demo.py MSFT / NVDA / TSLA --language koChatGPT subscription mode: run python -m cores.chatgpt_proxy.oauth_login once for browser-based auth, then run the orchestrator with PRISM_OPENAI_AUTH_MODE=chatgpt_oauth.
For production, docker compose up -d is recommended. Telegram distribution needs channel configuration; live trading needs Korea Investment & Securities API credentials. US PDF reports are saved to prism-us/pdf_reports/. Live performance is viewable at https://analysis.stocksimulation.kr/.
How does this agent compare with similar options?
The source names no direct competitors; it only positions itself as the AI assistant for investors (versus sponsor WrksAI's assistant for professionals). There is insufficient information for a factual comparison against other trading/analysis systems.