Polyseer: a research agent that reads a prediction market and argues about it

A market URL goes in, a forecast with sources comes out

The premise of Polyseer is that a prediction market tells you what might happen while the interesting question is why, and whether the price is right. You paste any Polymarket or Kalshi URL and the application reads the live market, which means the question, the prices, the close date and the resolution rules, and hands that to Valyu DeepResearch, described as an autonomous research agent that searches news, primary sources, data releases and related markets. The outputs are specific. A calibrated probability against the market price, with a plausible range and a confidence level. A side to take, buy YES, buy NO, or no edge. Evidence ranked by weight, for and against, each with its source. Base rate, catalysts, risks, and what would change the forecast. A full PDF report and a CSV of every factor affecting the market. And an email when the research finishes, so you can close the tab. The README states plainly that this is not financial advice, that the analysis is for entertainment and research, and that you should always do your own research.

A thin Next.js app around one external research job

The architecture section is blunt about what this is: Polyseer is a thin Next.js app around one external research job, and there is no in-process LLM pipeline. The README draws the whole flow as a diagram, and reading it explains the design. The user pastes a market URL and picks an effort level, which posts to the API. The server decides whether it is a Polymarket or a Kalshi market and fetches from the Gamma and CLOB APIs or the Kalshi trade API respectively, and both paths converge on the same set of market facts. From there the application builds a forecast prompt, a research strategy and an output schema, and hands them to a Valyu DeepResearch task. The task produces a structured forecast JSON, a PDF report with a CSV of factors, and a completion email, and the browser polls the status endpoint to turn the message trace into a live feed of searches, reads and sources. The important structural point is where the intelligence lives. The reasoning happens inside Valyu's hosted research agent, and this repository contributes the market reading, the prompt, the schema, the polling and the presentation.

Four request steps, from create to a stable forecast link

The request flow is described in four steps. Create is a POST to the API research route that validates the URL, fetches the live market, builds the prompt and creates a DeepResearch task with a JSON output schema, two deliverables, the PDF report and the CSV of factors, and an optional completion email, returning the task id. Watch is the browser navigating to the research page for that id and polling the status endpoint every few seconds, backing off to twenty seconds and pausing when the tab is hidden, with the message trace in the status response turned into an activity feed. Read is what happens when the task completes, where the structured output renders as a verdict card with the side, the probability against the market, the edge and the confidence, alongside the evidence, base rate, catalysts, risks and rationale, and the deliverable files download through a route that proxies the signed Valyu URL. History lists the account's DeepResearch tasks and keeps the ones this app created. The last detail is the one that matters for a tool you pay per run for: every forecast link is stable, so you can reopen it later to see progress or the finished report.

Self-hosted or Valyu mode, and who pays for the research

There are two operating modes, and the difference is mostly about authentication and billing. In self-hosted mode there is no authentication, research is billed to your own Valyu API key, and the research history is pulled straight from your Valyu account, which the README presents as ideal for personal use and development. In Valyu mode you sign in with Valyu using OAuth 2.1 with PKCE, and each user's research is billed to their own Valyu organisation credits through the platform proxy. The history follows the same split, tasks on your API key in one case and on the user's organisation in the other. The completion email address is set by DEEPRESEARCH_ALERT_EMAIL in self-hosted mode and by the signed-in user's email in Valyu mode, and the database is local SQLite in one case and your own Supabase project, holding users, in the other. The self-hosted mode is the recommended path in the environment template, and it is the one that requires no user accounts at all.

Four effort levels, each mapping to a research mode and a price

Research effort is chosen from the settings pill under the search box, with four levels, and each maps to a DeepResearch mode with a typical duration and a cost per run. Fast is the default and maps to the fast mode, taking about five minutes and costing 0.10 dollars per run. Standard maps to standard, taking ten to twenty minutes for 0.50 dollars. Deep maps to heavy, taking thirty to sixty minutes for 2.50 dollars. Max maps to max, running up to two hours for 15 dollars. The second deliverable file adds 0.10 dollars per run, and the environment template repeats the same range, from about 0.10 for the lowest tier to 15 for the highest. The README's framing is that deeper effort means more sources, more cross-checking and a longer report, which is a reasonable summary of what the extra money buys, and the pricing is worth reading before you leave a research running while you close the tab.

Self-hosting it takes a clone, pnpm and two environment variables

The self-hosted quick start is the shortest path to a running instance, and it is four commands plus an environment file:

git clone https://github.com/yorkeccak/polyseer.git
cd polyseer
pnpm install

# Create .env.local with:
# NEXT_PUBLIC_APP_MODE=self-hosted
# VALYU_API_KEY=valyu_xxx        # Get from platform.valyu.ai

pnpm dev

You then open localhost:3000, paste a Polymarket or Kalshi URL, pick a research effort and get the forecast. The environment template in the repository makes the same point in its own comments. Self-hosted mode needs only the app mode set to self-hosted and a server-side Valyu API key, with the optional DEEPRESEARCH_ALERT_EMAIL commented out, and the template warns you not to commit the local file. The prerequisites are Node.js 20 or newer and pnpm, plus a Valyu API key from platform.valyu.ai. One subtlety is worth knowing: completion emails deep-link to the research page under NEXT_PUBLIC_APP_URL, and that link has to be https, so on plain http during local development the email instead points at the Valyu platform's own report page.

The stack, the key files and the open licence question

The technology stack is Next.js 16 with React 19 and Turbopack, Tailwind CSS 4 with Radix UI and Framer Motion, TypeScript throughout, and Zustand for auth state. Storage is SQLite in self-hosted mode and Supabase in Valyu mode. The interesting dependency is the research layer itself, Valyu DeepResearch, which handles the research, the structured output, the deliverables and the completion emails, while the Polymarket and Kalshi APIs supply the live market data. The key files table shows a codebase organised by concern rather than by page. The prompt and the research strategy are built in one file, the JSON schema for the structured forecast and the parser that validates it in another, the activity feed conversion in a third, and the service layer handles create, status, list, cancel and deliverable lookup. The tools directory holds the Polymarket and Kalshi fetchers and URL parsing, and the market discovery module feeds the home page with trending markets and search. The project is private on npm with version 0.1.0, it has no GitHub releases, the last push was on 2026-10-14, and the repository reports no standard licence identifier, so if you intend to host this for more than yourself, the terms are something you have to establish rather than assume.

Editorial conclusion

Polyseer fits a researcher or trader who wants a written, sourced argument about a prediction market rather than a price, and who is prepared to pay per research run and to verify the citations before acting. It does not fit someone who wants a signal to trade automatically, since the output is a PDF and a CSV produced by an external agent, and the project itself says it is not financial advice. Before running it, decide between self-hosted mode and the hosted Valyu mode, read the effort table so the cost per run is explicit, and check the licence position, since the repository reports no standard licence identifier.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. yorkeccak/Polyseer on GitHub