# Oracle3

> Prediction-market research and execution tools with a Python API, JSON CLI, MCP server, and Agent Skills. MCP exposes public Kalshi/Polymarket data, fee-aware checks of supplied contract relations, and a local paper ledger. Live trading is a separate CLI capability.

## Start here

- [Agent quickstart](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/agent-quickstart/): client setup, deterministic first tool call, result interpretation, live research, and skill setup
- [MCP reference](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/mcp/): tool names, inputs, transports, and side effects
- [Agent contract regression](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/agent-harness/): frozen tool/executor cases, JSON evidence, and migration notes; does not evaluate LLM or Skill efficacy
- [Documentation home](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/): capabilities, examples, and current development scope

## Task and capability contracts

- Find candidates with `search_markets`; inspect `get_market`, `get_quote`, and `get_orderbook`. MCP venues are `kalshi` and `polymarket`. Prices are dollars per contract; sizes are contracts. Kalshi market IDs are tickers; Polymarket IDs are Gamma market IDs, not token IDs.
- Use `check_constraint` for supplied quotes without network access; use `check_constraint_live` for public live quotes and venue fee schedules. Relation labels are `implication`, `exclusivity`, `complement`, `same_event`, and `event_sum`.
- A result is conditional on the caller's relation and full fills at the quoted prices. The checker does not verify settlement rules or executable depth. On current main, `assessment` explicitly marks relation and execution as unverified. A positive gross violation can have a negative edge after fees.
- `get_market` returns Kalshi `rules_primary` and `rules_secondary`, and Polymarket `rules` and `resolution_source`, when available from the venue. If these fields are missing or ambiguous, check the venue's contract rules before treating markets as equivalent.
- `paper_order` buys against a live-book snapshot and writes a local ledger. Quantity must be finite and positive; limit price finite and strictly between 0 and 1. Invalid order inputs fail before venue HTTP. The ledger carries positions at cost and does not track resolution or P&L. No MCP tool places real orders.
- For safe paper retries, supply optional nonblank `request_id`. Once a result is saved, the same key and identical venue/market/side/quantity/limit return it; changed intent is an error. Unfilled/rejected results are remembered too. Omit the key only when each call should be a new order. `paper_reset` with `{"confirm":true}` clears positions, fills and retry history; `paper_portfolio` takes `{}`.
- Ledger operations lock across cooperating processes on a local filesystem and write atomically. A shared path is one paper account, not separate user state. Valid older JSON loads automatically. Corrupt/nonfinite/inconsistent ledgers fail closed and remain intact, even on reset: preserve a copy, restore a valid backup or set a fresh `ORACLE3_MCP_LEDGER` path and restart. Do not edit the file or remove its `.lock` sidecar while servers use it.
- `fair_value` returns a probability implied by a chosen Wang-transform parameter, not independently verified event truth. Recorded or fitted relations also require interpretation.

## Install and source

`pip install oracle3` and `uvx --from oracle3 oracle3-mcp` use the published package, which can lag main. The new assessment fields, paper retry handling and regression runner are documented for a main checkout; follow the agent quickstart to install that checkout and configure its absolute `oracle3-mcp` executable. Package installation alone does not install skills into a client.

- [Repository](https://github.com/YichengYang-Ethan/oracle3-prediction-market-agent)
- [README](https://raw.githubusercontent.com/YichengYang-Ethan/oracle3-prediction-market-agent/main/README.md)
- [AGENTS.md](https://raw.githubusercontent.com/YichengYang-Ethan/oracle3-prediction-market-agent/main/AGENTS.md): setup, checks and safety rules for coding agents
- [Agent Skills](https://github.com/YichengYang-Ethan/oracle3-prediction-market-agent/tree/main/skills): six workflows; setup is documented in the agent quickstart
- [Registry manifest](https://raw.githubusercontent.com/YichengYang-Ethan/oracle3-prediction-market-agent/main/server.json): package installation metadata
- [PyPI package](https://pypi.org/project/oracle3/)
- [CLI quickstart](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/CLI_QUICK_START/)

## Research and extensions

- [Fee frontier](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/research/fee-frontier/): calculations under stated schedules, not an observed trading return
- [Pricing working paper](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6468338) and [replication package](https://github.com/YichengYang-Ethan/prediction-market-pricing): model assumptions and estimates
- [oracle3-extras](https://github.com/YichengYang-Ethan/oracle3-extras): experimental venue, data, and wallet integrations; its venue coverage does not imply equivalent MCP support in core Oracle3
- [Open problems](https://github.com/YichengYang-Ethan/oracle3-prediction-market-agent/issues?q=is%3Aissue+label%3Aopen-problem)
- [Discussions](https://github.com/YichengYang-Ethan/oracle3-prediction-market-agent/discussions)

## Project background

- [Project overview and contributions](https://yichengyang-ethan.github.io/oracle3-prediction-market-agent/research-review/): Yicheng Yang's work at U Lab, Oracle3's extensions to the Coinjure foundation, MCP tools, reusable Skills, and directions for agent evaluation.