Prediction markets aggregate collective intelligence about future events, but extracting actionable insight from them remains difficult. Traders and researchers face information overload: academic papers, news cycles, market microstructure, and social sentiment all interact in complex ways. Manual analysis cannot keep pace with real-time market movements, yet most automated tools surface correlations without causal explanation. yorkeccak/Polyseer addresses this gap with a multi-agent AI architecture that conducts systematic bilateral research and aggregates evidence using Bayesian probability mathematics. Built in TypeScript with 665 GitHub stars and 121 forks, it targets developers who want rigorous, transparent analysis of Polymarket and Kalshi markets rather than opaque "vibes-based" predictions.
What is yorkeccak/Polyseer?
yorkeccak/Polyseer is an open-source prediction market analyzer maintained by yorkeccak. The project sits at the intersection of three technical domains: multi-agent AI systems, information retrieval, and probabilistic reasoning. Its last commit was February 27, 2026, indicating active development.
The tool's core premise is that prediction markets tell you what might happen, but not why. Polyseer fills this explanatory gap by deploying specialized AI agents that research both sides of a market question, classify evidence by quality, and mathematically aggregate findings into calibrated probability estimates. This design explicitly avoids confirmation bias through mandatory bilateral research and produces two probability outputs: a neutral objective assessment and a market-informed estimate.
The project is built for technical users—backend developers, quantitative researchers, ML practitioners—who can inspect and modify its reasoning pipeline. It is not a black-box consumer application. The self-hosted mode requires only three environment variables and runs locally with SQLite, making it accessible for personal use and development without external dependencies beyond API keys for Valyu search and OpenAI language models.
Technically, Polyseer is a Next.js↗ Bright Coding Blog 16 application using React↗ Bright Coding Blog 19, Tailwind CSS↗ Bright Coding Blog 4, and the AI SDK for LLM orchestration. It integrates GPT-4o and GPT-5 for reasoning, Valyu's search network for evidence retrieval, and platform-specific APIs for Polymarket and Kalshi market data. The architecture reflects current best practices in agentic AI systems: specialized components with defined interfaces, orchestrated by a central controller with feedback loops for quality improvement.
Key Features
Multi-Agent Research Architecture Polyseer implements five distinct agent types with separated concerns: a Planner that decomposes market questions into research pathways; Researcher agents that gather PRO and CON evidence in parallel; a Critic that identifies gaps and recommends follow-up searches; an Analyst that performs Bayesian probability aggregation; and a Reporter that generates human-readable output. This separation prevents the single-model failure mode where reasoning and evidence retrieval contaminate each other.
Bilateral Evidence Collection The system mandates research on both sides of any market question. Researcher agents execute parallel searches for supporting and contradicting evidence using Valyu Deep Search and Web Search. This design explicitly targets confirmation bias, a documented failure mode in both human and automated analysis systems.
Evidence Quality Classification All evidence receives a four-tier classification with explicit caps on influence: Type A primary sources (official documents, regulatory filings) capped at 2.0; Type B high-quality secondary (Reuters, Bloomberg, expert analysis) at 1.6; Type C standard secondary at 0.8; and Type D weak or speculative sources at 0.3. This quantified skepticism prevents low-quality information from dominating conclusions.
Bayesian Probability Aggregation The Analyst agent updates probabilities using log likelihood ratios with explicit correlation adjustments and cluster analysis. The system outputs two calibrated estimates: pNeutral (objective assessment) and pAware (market-informed). Mathematical foundations include LLR = log(P(evidence|YES) / P(evidence|NO)) with correlation-aware weighting.
Platform Integration Native clients for Polymarket and Kalshi APIs with unified market data representation. Users paste any market URL and receive structured analysis without manual data extraction.
Flexible Deployment Self-hosted mode with local SQLite (no authentication, unlimited queries with personal API keys) or Valyu mode with Supabase-backed OAuth for multi-user deployments.
Use Cases
Quantitative Research Workflow Researchers studying prediction market efficiency can use Polyseer to systematically document the information environment around specific markets. The evidence classification and quality scoring provide reproducible inputs for academic analysis of how information incorporation varies across market types.
Pre-Trade Due Diligence Traders with existing positions or considering entry can use Polyseer to surface factors they may have missed. The bilateral research design specifically catches contrarian evidence that confirmation-prone manual research would overlook. The pNeutral versus pAware divergence can flag when market prices appear misaligned with fundamental evidence.
Event-Driven Analysis Journalists and policy analysts tracking specific outcomes (election results, regulatory decisions, economic indicators) can use Polyseer to maintain structured evidence bases. The automated research pipeline scales to monitoring multiple markets simultaneously, with the Critic agent flagging when new information requires updated assessments.
Multi-Agent System Education Developers learning to build agentic AI systems can study Polyseer's architecture as a production reference implementation. The explicit agent boundaries, feedback loops between Critic and Researcher, and mathematical aggregation layer demonstrate patterns applicable beyond financial analysis.
Custom Research Pipelines The modular agent design allows extension to other information domains. While currently focused on prediction markets, the Planner-Researcher-Critic-Analyst-Reporter pattern applies to any domain requiring systematic evidence synthesis with uncertainty quantification.
Installation & Setup
Polyseer's self-hosted mode requires minimal configuration. The README provides exact commands that should be reproduced precisely:
# Clone the repository
git clone https://github.com/yorkeccak/polyseer.git
cd polyseer
# Install dependencies
npm install
Create .env.local with exactly three variables:
# Application mode
NEXT_PUBLIC_APP_MODE=self-hosted
# Search API access (obtain at platform.valyu.ai)
VALYU_API_KEY=valyu_xxx
# Language model access (obtain at platform.openai.com)
OPENAI_API_KEY=sk-xxx
Then start the development server:
npm run dev
The application serves at localhost:3000. No additional database setup is required—SQLite initializes automatically. This minimal footprint makes Polyseer viable for personal workstations without container orchestration or managed database services.
For production or multi-user deployments, Valyu mode requires additional OAuth and Supabase configuration. The README notes this is currently restricted: "Valyu OAuth apps will be in general availability soon. Currently client id/secret are not publicly available. Contact contact@valyu.ai if you need access." This limitation should be noted when planning organizational deployments.
Real Code Examples
The README documents the Planner agent's output interface, which defines the contract between planning and research phases:
interface Plan {
subclaims: string[]; // Causal pathways to outcome
keyVariables: string[]; // Leading indicators to monitor
searchSeeds: string[]; // Targeted search queries
decisionCriteria: string[]; // Evidence evaluation criteria
}
This interface demonstrates Polyseer's structured approach to decomposing market questions. The subclaims field captures distinct causal mechanisms, preventing the common failure mode where analysis conflates correlated factors. keyVariables enables monitoring of leading indicators as new information arrives. searchSeeds provides explicit query generation rather than relying on model improvisation. decisionCriteria establishes evaluation standards before evidence collection, reducing hindsight bias in assessment.
The environment configuration for self-hosted mode is the primary setup code in the README:
# ===========================================
# Self-Hosted Mode Configuration
# ===========================================
NEXT_PUBLIC_APP_MODE=self-hosted
NEXT_PUBLIC_APP_URL=http://localhost:3000
# ===========================================
# Required API Keys
# ===========================================
# Get your Valyu API key at: https://platform.valyu.ai
VALYU_API_KEY=valyu_your_api_key_here
# Get your OpenAI API key at: https://platform.openai.com
OPENAI_API_KEY=sk-your_openai_api_key_here
This configuration reflects Polyseer's operational simplicity in self-hosted mode. The NEXT_PUBLIC_APP_MODE switch determines database backend (SQLite versus Supabase) and authentication requirements. The two API keys represent the only external dependencies—no additional vector database, message queue, or inference hosting is required.
The README does not contain additional code examples beyond these configuration snippets and the TypeScript interface. Developers should expect to inspect source code directly for implementation details of agent behaviors and mathematical aggregation.
Advanced Usage & Best Practices
API Key Management Self-hosted mode uses your personal API keys without rate-limiting intermediaries. Monitor usage directly through Valyu and OpenAI dashboards. For high-frequency analysis, consider implementing request batching or caching at the application layer—Polyseer does not currently include these optimizations.
Evidence Interpretation The pNeutral and pAware divergence deserves attention. Large gaps suggest market inefficiency or information asymmetry; convergence indicates well-incorporated public information. Track this divergence over time rather than treating single-point estimates as definitive.
Extending Agent Behaviors The modular agent architecture supports modification. Developers can adjust evidence type caps, add custom search seeds, or implement domain-specific classification rules. Maintain the bilateral research invariant—unilateral modification risks reintroducing confirmation bias.
Database Considerations Self-hosted SQLite suffices for individual use. For concurrent access or historical analysis across many markets, migrate to the Supabase-backed Valyu mode when generally available. The README notes current restrictions on OAuth credentials for this path.
Model Selection Polyseer targets GPT-4o and GPT-5. While the AI SDK abstraction permits model swapping, the reasoning quality of planning and analysis phases depends on capable foundation models. Downgrading to less capable models likely degrades evidence synthesis quality disproportionately.
Comparison with Alternatives
| Tool | Approach | Key Difference | Trade-off |
|---|---|---|---|
| Polyseer | Multi-agent with Bayesian aggregation | Explicit bilateral research, evidence quality classification | Requires self-hosting or waitlist for managed mode |
| Manifold Markets | Social prediction with discussion | Native market creation, community deliberation | Less systematic evidence retrieval, no automated aggregation |
| Metaculus | Crowd forecasting with commentaries | Large forecaster base, track record scoring | Human-latency updates, no real-time automated research |
Polyseer's distinctive value is automated systematic research with mathematical aggregation, versus the social deliberation models of alternatives. This suits users who need scalable, reproducible analysis rather than community consensus. The trade-off is operational complexity: Polyseer requires API keys and hosting, while alternatives offer fully managed access.
FAQ
What prediction markets does Polyseer support? Polymarket and Kalshi, with unified data representation via platform-specific API clients.
Is a license specified? The README states MIT License, though the repo_stats list "Not specified." Check the repository directly for current licensing.
Can I run Polyseer without OpenAI or Valyu accounts? No. Both API keys are required for core functionality—Valyu for search, OpenAI for agent reasoning.
What database does self-hosted mode use? Local SQLite, automatically created. No manual schema management required.
How current is the evidence? The Valyu integration provides real-time search across academic papers, web sources, and market data—not stale training data.
Is this financial advice? Explicitly not. The README states: "Polyseer provides analysis for entertainment and research purposes only."
What Node.js version is required? Node.js 18 or higher, per the README prerequisites.
Conclusion
yorkeccak/Polyseer occupies a specific niche: developers and researchers who want transparent, mathematically grounded analysis of prediction markets without black-box opacity. Its 665 stars and active commit history suggest growing interest in reproducible AI-assisted research tools. The multi-agent architecture with explicit bilateral research and Bayesian aggregation represents a principled approach to a problem domain often addressed with less rigor.
The tool is best suited for technical users comfortable with self-hosting, API key management, and TypeScript codebases. It is not a turnkey consumer application. Researchers studying market efficiency, traders seeking systematic due diligence, and developers building agentic AI systems will find the most value.
The current limitation is managed deployment availability—Valyu mode remains restricted. For individual use, the self-hosted mode removes this barrier entirely.
Ready to explore systematic prediction market analysis? Clone yorkeccak/Polyseer on GitHub and start with the three-variable self-hosted setup.
