PRISM-INSIGHT: A Multi-Agent Stock Analysis Stack Bound to the Korean Market
AI-based stock analysis and trading system
At a glance
- What is it?
- PRISM-INSIGHT wires more than a dozen LLM agents into a pipeline that scans for surge stocks, writes reports, and places trades. The README is explicit about its market coverage and its AGPL-3.0 licence, and both shape who can realistically run it.
- Who is it for?
- Adopt PRISM-INSIGHT if you already hold a ChatGPT Plus or Pro subscription, you trade KRX or US equities, and you are willing to publish modifications under AGPL-3.0. Skip it if you need a strategy with a verified multi-year track record, if you trade markets outside KRX and US, or if you cannot accept that a bad LLM call can reach a live brokerage order.
- Can I use it commercially?
- Yes, with strict conditions. AGPL-3.0 is a network copyleft licence: if people use a modified version over a network, for example as a hosted service, you must offer them its source code under the same licence.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Problem PRISM-INSIGHT Targets: Surge Detection Plus Report Generation
Most retail screening tools stop at a filter. They return a list of tickers that match a price or volume condition, and the interpretation is left to you. PRISM-INSIGHT takes the next step: it runs a set of specialized agents over the candidates, produces a written report, and then executes trades. The README describes the goal as detecting surge stocks, generating analyst-grade reports, and executing trades automatically, with 13 or more specialized AI agents collaborating on that chain.
The intended user is someone who already trades Korean or US equities and wants the screening-to-order path compressed into one system. The project ships a mobile app on both Google Play and the App Store, a Telegram alert channel in English and Korean, and a live dashboard at analysis.stocksimulation.kr. That distribution footprint suggests the author is targeting individual traders rather than institutions. There is no mention of a multi-tenant deployment model, no role-based access description, and no audit trail format documented in the material supplied. If you need those, this is not the project for you.
How the Agent Pipeline Is Wired Together
The repository is Python, requires 3.10 or higher, and is organized around an orchestrator entry point. The README shows the invocation as stock_analysis_orchestrator.py with a --mode flag, and the example passes morning. That implies at least one other mode exists, most likely an evening or closing pass, though the supplied material only confirms morning.
The agent layer sits behind that orchestrator. The README states that 13 or more specialized agents collaborate, and names the model providers it targets: OpenAI GPT-5, Anthropic Claude Sonnet 5, and ChatGPT Plus or Pro through a Codex OAuth proxy. Those are separate authentication paths into the same analysis pipeline, which means the model choice is a configuration concern rather than a code fork. The release notes for v2.19.0 mention a redesign of the trade execution structure and a report SDK-neutralization step, which reads as an effort to decouple report generation from whichever model provider is active.
One detail worth flagging: the README does not document the message format between agents, the number of turns per analysis, or how the orchestrator handles a partial failure when one agent returns nothing usable. For a system that ends in order placement, that gap matters more than the agent count.
Running It With a ChatGPT Subscription Instead of an API Key
The most concrete operational detail in the README is the Codex OAuth path. It lets you run analysis against a ChatGPT Plus or Pro subscription rather than a metered API key. The commands are given directly:
python -m cores.chatgpt_proxy.oauth_login
python -m cores.chatgpt_proxy.oauth_login --force
PRISM_OPENAI_AUTH_MODE=chatgpt_oauth python stock_analysis_orchestrator.py --mode morning
The first command opens a browser for ChatGPT authentication. The --force variant re-authenticates, which the README frames as the path for switching accounts or refreshing expired tokens. The environment variable PRISM_OPENAI_AUTH_MODE is what switches the orchestrator onto that credential source, and the README notes tokens refresh in the background so re-login is only needed on an account or password change.
That is a real cost lever. A subscription is a flat monthly fee; API usage is metered per token. For a pipeline that runs many agents over many candidates, the difference in monthly spend is not small. The trade-off is that you are routing automated analysis through a consumer subscription endpoint, and the README does not state whether that usage pattern is within the provider's terms. Verify that yourself before pointing it at anything that matters.
Stance: A Public Record That Cannot Be Backfilled
Stance is the project's strategy leaderboard, and its design choice is the interesting part. Records start at registration. The README states plainly that there are no uploaded track records and no backfills, and that the server seals decision time and price before calculating outcomes. That ordering is what makes the record hard to fake: the decision is timestamped before the result is known.
The leaderboard separates KRX and US rankings and, alongside return, shows worst drawdown, average invested exposure, and record rate. Ranking eligibility requires 63 trading days and 20 closed trades where each used at least 1 percent of assets. Until then a strategy sits under a building-a-record state. No live brokerage account or broker key is required to participate, which lowers the barrier considerably.
The limitation is equally clear: a new strategy has no history, and the 63-day plus 20-trade threshold means you will not see a ranking for months. Anyone who wants to evaluate a strategy on day one cannot use Stance for that. It is a forward-only instrument.
Where the Documentation Goes Quiet
Several things a trading system needs are absent from the supplied material. There is no description of the order-placement path itself: which broker adapter is used, whether orders are market or limit, what happens on a partial fill, or how the system behaves when the market is closed. The v2.19.0 release note mentions a full redesign of the trade execution structure, which suggests the author is actively working on this area, but the README does not document the result.
The risk controls are also undocumented. There is no stated maximum position size, no daily loss limit, no kill switch, and no dry-run mode described. For a system whose headline claim is automatic trade execution, that is the largest gap in the material. A bad model output reaching a live order is the failure mode that matters here, and nothing in the supplied text indicates what stops it.
Market coverage is another boundary. The material names KRX and US equities and the KIS brokerage integration in the v2.20.0 notes. If you trade European or Japanese equities, there is nothing here for you. The README also does not state the minimum capital, the supported account types, or whether short selling is handled.
Licence and Maintenance: AGPL-3.0 With a Fast Release Cadence
The project is AGPL-3.0. That is the strongest of the common copyleft licences, and its network clause is the part that catches people. If you modify PRISM-INSIGHT and let users interact with it over a network, the licence requires you to offer those users the corresponding source. Running it privately for your own trading does not trigger that. Running a modified version as a service for others does. This is not legal advice; read the licence text and talk to counsel if you plan to host it.
The release cadence is fast. The three most recent releases span roughly seven weeks: v2.19.0 on 2026-07-21, v2.20.0 on 2026-08-05, and v2.21.0 on 2026-08-26. Each carries substantive changes, including a trade execution redesign, a KRX and KIS quote feed switch, a KakaoTalk integration, and an observability addition. Upgrading a system that places orders is not a routine dependency bump. Every one of those releases touches either the execution path or the data feed, and both are places where a regression has financial consequences. Budget time to read the diff before you pull, and keep a pinned version you can roll back to.
The AGPL also interacts with the commercial layer. The mobile app, the credits system mentioned in the launch promotion, and the hosted dashboard are the author's, and the licence does not oblige them to give you the hosted service. You can run the code; you cannot assume you get the infrastructure.
How It Compares to a Plain Screening Script
The obvious alternative is a conventional screener plus a scheduled script: pull quotes from a data provider, apply price and volume filters, write the matches to a file or a chat message, and stop there. Tools in that category, including the screening features built into most retail brokerage platforms, do the detection step well and cost nothing extra.
The difference is what happens after detection. A screener gives you a ticker and a number. PRISM-INSIGHT gives you a ticker, a written report assembled by several agents, and an order. Whether that extra layer is worth it depends on how much you trust the report. A screener is deterministic: the same inputs produce the same list. An LLM pipeline is not, and the README does not describe any evaluation harness that would tell you how often the reports are useful. The Stance leaderboard is the closest thing to that, but it measures strategy returns rather than report quality, and it only accumulates forward.
If your actual need is a ranked list of candidates to review manually, a screener is cheaper and more predictable. PRISM-INSIGHT is for the case where you want the interpretation and the execution inside the same loop.
Editorial conclusion
Adopt PRISM-INSIGHT if you already hold a ChatGPT Plus or Pro subscription, you trade KRX or US equities, and you are willing to publish modifications under AGPL-3.0. Skip it if you need a strategy with a verified multi-year track record, if you trade markets outside KRX and US, or if you cannot accept that a bad LLM call can reach a live brokerage order. Before anything else, read the order-placement path in the trading execution code and confirm which broker adapter is wired to it.
Community notes