oddsrail

The rail AI agents use to trade prediction markets.

An MCP server that gives any agent (Claude Code, Claude Desktop, or anything MCP-compatible) prediction-market access across Polymarket and Kalshi: market search, orderbooks, price history, positions, and order routing, with on-chain builder-code attribution on Polymarket, plus two premium signal tools (in-play overshoot/fade detection, resolution dispute-risk).

Free to use, and free of fees. oddsrail ships with a project builder code registered at 0 bps, so orders routed through it are attributed without adding a single basis point to anyone's trade. The project's income is a share of Polymarket's weekly builder reward pool, paid by Polymarket's own program, not by you. Running your own builder profile instead is one environment variable (ODDSRAIL_BUILDER_CODE), and server_info always tells you which code is in use. No fee tiers, no paywalled tools, no account required.

How oddsrail compares

Verified against each alternative directly (their repos, live endpoints, and registry entries, September 2026), not from their marketing:

Verified 2026-09-02 from each project's own README or docs (pmxt: github.com/pmxt-dev/pmxt; Simmer: docs.simmer.markets; agent-skills: github.com/Polymarket/agent-skills). "Not documented" means exactly that, not "absent". Re-check before quoting; these projects move.

The wedge, in one line: pmxt is the reference for trading everywhere; Simmer is the reference for an agent economy with a sandbox and a reputation layer; oddsrail is the reference for trading correctly, non-custodially, with attribution you own.

Where the others are honestly ahead: pmxt trades three venues to our two and covers a dozen more for data, with hosted convenience and a community many times ours. Simmer has a virtual-balance sandbox, stop-loss and take-profit rails we do not have, a public reasoning/reputation layer, and a strategy-skills marketplace. Polymarket's agent-skills is the venue's own documentation and covers bridging and deposits, which oddsrail does not.

¹ The raw Polymarket API has the endpoints. It also models rejections as ok:false return values, orders its books worst-first, ships a trades endpoint that returns the market's public tape, and enforces an undocumented $1 minimum notional. oddsrail exists because we hit every one of those and encoded the fix.

² Kalshi prices are dollar strings (integer cents were removed 2026-03), its orderbook is bids-only on both sides, and its current SDK requires Python ≥3.13. All normalised here.

Quickstart

Python 3.11+ required.

pip install oddsrail

claude mcp add --transport stdio oddsrail -- oddsrail

Or from a clone, without installing:

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

claude mcp add --transport stdio oddsrail -- /abs/path/to/oddsrail/.venv/bin/python -m oddsrail.server

Then ask the agent: "search markets about the World Cup final and run the overshoot signal on the favorite".

Install in one step

The plugin and the one-click links launch the server with uvx, so they need uv on the machine. Without uv, pip install oddsrail gives you an oddsrail command to point any client at. Everything starts in dry-run.

The four skills (skills/*/SKILL.md) are generated from the server's own MCP prompts by scripts/gen_skills.py, and a test fails if they drift, so a skill and the prompt it mirrors can never disagree.

Hosted: nothing to install

mcp.oddsrail.app runs the same server as a remote MCP endpoint with accounts, so an agent inside Claude can use it without a machine of its own. Add the URL as a custom connector (Pro, Max, Team and Enterprise plans), sign in with your email when Claude asks, and every call from then on carries your account.

What the hosted server is, in one breath: Polymarket market data, the signal tools, check_order, and paper trading with a $1,000 virtual bankroll per account, filled against the live book. What it is not: a place where money moves. It holds no wallet keys, executes no real order, and serves no Kalshi tools (Kalshi's API Developer Agreement limits API use to a member's own trading, so a shared service cannot route it). Account-scoped tools such as open_orders and the gasless relayer tools are absent, because on a shared server they would describe nobody's account. Twenty-four tools remain: the public-data and paper tools plus arena_register, arena_unregister and arena_status, which put the account's paper ledger on the public board.

Live trading stays self-hosted: pip install oddsrail with your own key, and the same place_order posts real orders when you set ODDSRAIL_DRY_RUN=0. The paper ledger you build up in Claude is yours to reset with paper_reset; nothing else about the account exists. Privacy policy: oddsrail.app/privacy. Source: oddsrail/cloud/ and oddsrail/hosted.py; deployment notes in deploy/cloud/.

Builder page and arena

oddsrail.app/build is a set of switches: strategies (fade overshoots, buy near-certain resolutions, trade your own probability, follow the move, two-sided quotes), risk rules (stop loss as a review rule, take profit, daily loss limit, never add to losers, exposure caps) and market hygiene (dispute risk, fill quality, watching the book). Press Run and the switches execute on the hosted server as a deterministic paper pass (oddsrail/cloud/runner.py): the universe is scanned, each strategy's signal is computed with the same code the MCP tools use, every order goes through check_order, fills come from the live book into a paper ledger that belongs to your browser, and the pass comes back as a list of decisions with verdicts and reasons. No account, no install, no key, no model. One pass is one tick at the current book, not a backtest: the ledger carries positions and resting quotes between passes, and nothing is watched in between, so a record needs many passes. The same switches also compose into a prompt for running the pieces inside your own Claude with the connector; the two-sided-quotes piece carries the maintainer's own warning about markout either way.

oddsrail.app/arena is the public board. Paper division: hosted accounts that entered themselves with arena_register, ranked by return on the virtual bankroll, served as JSON from mcp.oddsrail.app/arena/paper.json. Live division: wallets whose Polymarket fills carry the oddsrail builder code, ranked by attributed volume from the public feed, with realized and open P&L from Polymarket's data API; register by pull request or the form on the page (site/arena/agents.json). No prizes yet; the rules are on the page.

See the footguns yourself, no keys

pip install oddsrail && python examples/footguns.py

Six things the venue APIs get wrong, shown live against the public endpoints: the worst-first order book, Kalshi's bids-only dollar-string ladders, geoblocks that look healthy until the order, the advisory geoblock verdict, the deprecated Gamma endpoint, and naive cross-venue matching. Each section prints the raw trap, then the oddsrail call that returns the right answer. About ten seconds, read-only.

How attribution works (CLOB V2, verified Aug 2026)

1. Get your builder code (a bytes32) at polymarket.com → Settings → Builders. Set your fee rates there: taker up to 100 bps, maker up to 50 bps, additive on top of platform fees, settled to your builder wallet.
2. export ODDSRAIL_BUILDER_CODE=0x... where the server runs.
3. Every order any agent routes through place_order has the code placed in the V2 order struct's builder field before signing, so attribution is on-chain, visible in every OrderFilled event on CTF Exchange V2.
4. Verify with the builder_stats tool (public builder-trades endpoint + leaderboard).

If you skip this, orders carry the bundled oddsrail builder code (0xa576c5ce…, registered at 0 bps maker / 0 bps taker), costing you nothing and funding the project. If you set your own, yours wins; the default is a default, not a lock-in.

The oddsrail builder profile is Verified in Polymarket's builder program (2026-09-02), and Polymarket's builder team confirmed builder-code attribution as the right pattern for a self-hosted, non-custodial tool: no keys ship with the product, and the code is attached and signed by the operator's own wallet.

Environment variables

Status

Offline tests: 175 tests covering the paths where a bug costs money: the Kalshi yes/no→bid/ask translation, Kelly sizing, book walking, cross-venue pairing, signal edge cases, the dry-run safety net, and jurisdiction-failure handling (a geoblock must never read as an empty search result or a resting order), plus the hosted server end to end (dynamic client registration, PKCE, magic-link sign-in, token rotation, one paper ledger per account) against a real local process. They need no keys and no network:

pip install -e ".[dev]" && pytest

CI runs them on Python 3.11, 3.12 and 3.13 on every push and pull request.

Live venues: every read tool and both dry-run order paths have been driven end-to-end against real Polymarket and Kalshi through a real MCP client session. Two attributed Polymarket orders (a buy and a sell) have been placed and confirmed on-chain.

Not yet exercised live: the Kalshi order placement path. Its request shape is unit-tested and its endpoint verified, but no order has been sent to a real Kalshi account, because the author does not yet have a funded, verified one. This is untested, not untestable. Treat kalshi_place_order as unproven and start in dry-run.

Where this works

Two different things can stop oddsrail from trading, and they have opposite remedies. One is a venue restriction, enforced at the order. The other is a network filter, which breaks the connection itself.

Polymarket restrictions. Polymarket publishes its restricted-jurisdiction list as an API reference: https://docs.polymarket.com/api-reference/geoblock. There are three tiers. OFAC-sanctioned jurisdictions (Iran, Syria, Cuba, North Korea, and the Crimea, Donetsk and Luhansk regions of Ukraine) are blocked on both the frontend and the API, with no new orders and no closing of existing positions. A longer second tier is close-only on both the frontend and the API: existing positions can be closed, new ones cannot be opened. It includes the United States, the United Kingdom, France, Germany, Italy, Poland, Slovakia, Belgium, Singapore, Australia, New Zealand, Brazil, Russia, Taiwan, Thailand and the Canadian provinces of Ontario, Quebec, British Columbia and Alberta. A third group, Ireland, Japan, Malta (sports only) and the Netherlands, is close-only on Polymarket's frontend, with the API explicitly not restricted.

Note the shape of that failure: it lands on the order, not the connection. Public reads answer normally, so oddsrail will look like it is working right up until an order is rejected. Verified against Polymarket's documentation on 2026-08-31; Polymarket updates the list without notice, so read the URL rather than this paragraph.

Kalshi restrictions. Kalshi is a single CFTC-designated contract market and it does admit members outside the United States, but its Member Agreement §VI names a long list of Restricted Jurisdictions whose members may not trade Event Contracts, among them Australia, Belgium, Canada, France, Ireland, Italy, New Zealand, Poland, Portugal, Singapore, Switzerland, the United Kingdom, Hungary, India, the United Arab Emirates and mainland China. The list is published in Kalshi's Exchange Notice of 22 June 2026 (https://kalshi-public-docs.s3.amazonaws.com/regulatory/notices/Kalshi%20Exchange%20Notice%20(Updated%20Member%20Agreement)%20(22%20June%202026).pdf), and Kalshi reserves the right to change it. The same section is explicit that the restriction applies only to trading Event Contracts and does not by itself bar membership or non-trading access, so oddsrail's Kalshi read tools stay usable even where its order tools do not.

The two lists overlap heavily. Kalshi is not a general fallback for a Polymarket-restricted operator, and the difference runs in both directions. Among the jurisdictions polymarket.com lists as close-only, Germany, Brazil, Slovakia and the United States are not on Kalshi's restricted list; Japan and the Netherlands are restricted by neither API (only by Polymarket's frontend). Check both lists for your own jurisdiction rather than assuming the other venue is open. The US case has its own wrinkle.

The United States. polymarket.com, the venue oddsrail talks to, is close-only for the US. Polymarket separately operates Polymarket US (polymarket.us), run by QCX LLC as a CFTC-regulated Designated Contract Market. oddsrail does not support it. It is a different API host, a different authentication model (API-key headers rather than EIP-712 wallet signatures), a different SDK and a different funding rail. A polymarket.us account and its keys will not work with this server. Kalshi does not list the US as restricted, so for a US operator Kalshi is the venue oddsrail can actually reach, with no builder-code attribution, since Kalshi's REST API has no such field.

Network filters. Separately from any venue rule, a national filter can block the domains outright. Turkey does this: Polymarket does not restrict Turkey, and Turkey is not on Kalshi's list either, but Turkish ISPs block polymarket.com. That is a connectivity problem, not an eligibility one, and it looks different: DNS failures, TLS errors, resets, or an ISP interstitial page served where JSON was expected. oddsrail classifies both shapes and tells the calling agent which one it hit.

Eligibility is the operator's, not the tool's. oddsrail is self-hosted and non-custodial, which is a real advantage and also means you hold the account and you make the venue's representations; there is no intermediary making them for you. Polymarket's trading flow requires an attestation that you are not a U.S. person, are not located in a restricted jurisdiction, and are not "using a VPN or other measures to circumvent or attempt to circumvent" restrictions, and states that Polymarket reserves the right to put a non-compliant wallet in close-only mode. Kalshi's §VI is a representation about where you are domiciled, organized and located, re-made each time you place an order. server_info reports Polymarket's geoblock verdict for this machine's IP, but a technical probe is not a compliance check: the terms bind on residence, citizenship and incorporation, not on egress IP. Read the terms; if any of this matters to you, get your own legal advice. Nothing here is legal advice.

The signal logic, the MCP layer and the whole test suite run fine offline regardless.

Guardrails: limits the agent cannot argue with

Anyone handing keys to an agent wants three things first: a cap on one order, a cap on a session, and a fence around which markets it may touch. All three are operator-set environment variables (table above), enforced before any request goes out, in dry-run as well as live, so the agent meets the fence in rehearsal. A refusal is a structured answer that names the rule, the limit and the request:

{"accepted": false, "blocked_by": "guardrail", "rule": "max_order_notional",
 "limit": 25.0, "requested": 99.5, "note": "refused by an operator-set guardrail ... Nothing was sent."}

The session counter lives in the server process; restarting it resets the budget, which is the operator's call. server_info reports the active limits and how much of the session budget is used.

Paper trading: dry-run with a memory

By default, every dry-run Polymarket order is filled against the live order book, walked within the limit price; whatever does not fill rests as a paper order and fills later if the market crosses it. paper_positions reports cash, positions at current marks, realized and unrealized P&L and the resting paper orders; paper_reset starts over. The ledger is one local JSON file. Be clear about what this is: fills assume no queue position, no latency, no market impact and no fees, so paper results are an upper bound on the same strategy live. Kalshi dry-run orders still return the intent only.

Realtime: watch the book move

watch_book(token_id, seconds, max_events) subscribes to a token's realtime stream and returns the events that arrived (book snapshot, then price changes and trades), bounded to at most 60 seconds so an agent cannot hang a session on a quiet market. Use it after get_orderbook when the decision depends on the book moving, not just where it is.

If the stream fails with CERTIFICATE_VERIFY_FAILED while the REST tools work, your Python has no CA bundle (common with python.org macOS installs). oddsrail classifies that as local_tls and tells the agent the fix: run Install Certificates.command from the Python folder in /Applications, or set SSL_CERT_FILE to the path printed by python -m certifi.

Gasless position management (relayer)

Three tools move collateral without paying gas, through Polymarket's relayer: split_position (USDC → a full YES+NO set), merge_positions (matching YES+NO → USDC, or max), and redeem_positions (a resolved market's winning shares → USDC). All three respect dry-run and return the relayer transaction id and hash plus the terminal outcome.

They use your own Relayer API key, created at polymarket.com → Settings → Relayer API keys and exported as POLYMARKET_RELAYER_API_KEY + POLYMARKET_RELAYER_API_KEY_ADDRESS. That is the pattern Polymarket's builder team recommends for a self-hosted tool: no builder secret ships with oddsrail, and each operator authenticates the relayer as themselves. Relayer limits are per builder tier: 100 requests/day unverified, 10,000 verified. Without the key the tools return a structured "not configured" answer and send nothing; they never fall back to a gas-paying broadcast from the signer.

Exercised live (2026-09-02): a 1 USDC split and the matching merge went through the relayer from this code, gasless, on the maintainer's test account with its own Relayer API key. Relayer ids and Polygon transaction hashes are in docs/live-proof.md. redeem_positions is still unproven live: it needs a resolved market with winning shares, which that account has not held yet. redeemable_positions lists what the configured wallet could redeem or merge right now, and the settle_resolved prompt chains the two.

Kalshi (venue #2)

Kalshi is bring-your-own-key and single-tenant by design: the operator supplies their own API key, trades their own account, and this server caches nothing. That is deliberate: Kalshi's Developer Agreement limits API use to a member's own trading (§3), bars facilitating other members' trading (§3.2) and sublicensing the API (§3.7), and restricts storing/sharing API data (§3.1). A hosted multi-tenant Kalshi service would not be compliant; a self-hosted one is.

Attribution does not exist here. Kalshi Builder Codes are a Solana/DFlow/Jupiter integration; there is no builder or affiliate field anywhere on the REST API, so Kalshi order flow cannot be attributed or monetised the way Polymarket's can. Kalshi is in oddsrail for coverage and signal reach, not for routing revenue.

Two shapes on this API are easy to get wrong, so oddsrail normalises both:

• Prices are dollar strings, not cents ("0.5600"), sizes are fixed-point strings ("10.00"); the legacy integer-cent fields were removed in March 2026. All arithmetic uses Decimal.
• The orderbook is bids-only on both sides. yes_dollars and no_dollars are both bid ladders, ascending, so the best bid is the last element, and a NO bid at $0.99 is a YES ask at $0.01. kalshi_get_orderbook returns a conventional best-first bid/ask view of the YES book plus the raw ladders.

Order placement speaks natural terms, outcome (yes/no), action (buy/sell), price = probability of that outcome, and translates to Kalshi's YES-book bid/ask internally (buy NO @ 0.25 becomes ask @ 0.75). That translation is exhaustively unit-tested (tests/test_money_paths.py), since it is the obvious place to ship an inverted-position bug.

Credentials: KALSHI_KEY_ID plus KALSHI_PRIVATE_KEY_PATH (PKCS#8 PEM) or KALSHI_PRIVATE_KEY. Set KALSHI_DEMO=1 to hit the demo environment. Read tools need no key at all.

Cross-venue tools

• find_markets(query): searches Polymarket and Kalshi in one call and returns one normalised shape per market: venue, market_id (the id that venue's order tool takes), title, yes/no price as probabilities in (0,1), best bid/ask, spread, 24h volume, close time, and trade_with naming the tool to call. Use this when you do not already know the venue.
• quote_cost(venue, market_id, side, size): what a size would actually cost, by walking the book rather than reading the top level. Returns average fill price, slippage vs best, notional, levels consumed, and whether the size is fillable at all, plus Polymarket's per-market fee schedule where it publishes one. Kalshi does not publish fees in its market payload, so they are reported as unknown rather than estimated.
• compare_venues(query): candidate same-event listings across venues. Not an arbitrage scanner. Matching an event across venues is an unsolved entity-resolution problem: naive title overlap cheerfully pairs a Brazilian election with a Ukrainian one and reports a 70-point "gap" that is fiction. Two gates apply (title similarity ≥ 0.5 and close dates within a week), so it usually returns nothing, which is the honest answer. A price delta between candidates is reported as yes_price_difference, never as profit.

Kalshi search

Kalshi has no text-search endpoint. oddsrail searches by event (the human-readable index, with with_nested_markets) rather than paging tens of thousands of machine-named markets, and matches on word boundaries, without that, "fed" matches "German Bundestag" and a Fed-rate query returns German election markets. Results carry truncated, because a bounded scan means an empty result is not proof a market does not exist.

Order lifecycle & discovery

• order_status(order_id): resting / partially_filled / filled / gone, with size_matched. The answer an agent needs after place_order.
• my_fills(), my_positions(): the operator's executions and holdings, no address juggling. (Fills come from the Data API activity feed; the SDK's list_account_trades returns the market's public tape and is not used.)
• cancel_all_orders(): kill switch, flattens every resting order at once.
• resolution_criteria(venue, market_id) returns the full resolution contract: what resolves YES, who resolves it, from which sources. Read it before trusting a price.
• closing_soon(hours): markets closing within N hours on either venue, where activity concentrates.

Workflow prompts

MCP prompts show up in clients as ready-made workflows, and they encode the order of operations that keeps an agent out of trouble; the sequencing is the expertise, which a flat tool list cannot convey.

• /find_fade_setup(query, bankroll): signal → book → cost → resolution → size → dry-run, with the rejection criteria at each step
• /check_cross_venue_edge(query): candidates → settlement audit → cost on both legs, and says plainly when the answer is "no edge"
• /daily_review: positions, resting orders, fills, closing-soon, attribution

Risk & settlement

• settlement_audit(polymarket_id, kalshi_ticker): the check that decides whether a cross-venue price difference is an edge or a mismatch. Compares close times, resolution sources, UMA dispute status and market structure on live data with no pre-curated pair list, returning ok / caution / block with reasons, and listing the checks it did not perform.
• check_order(venue, market_id, side, price, size, intent): the last step before place_order. Deterministic checks of the proposed order against the operator's own words and the live market: does the market exist and accept orders, do the intent's words match the market and the YES/NO side, is the price sane against the book, is the size above Polymarket's $1 minimum and inside the guardrails, is there liquidity within the limit, is a resolution source named. Returns ok / caution / block with the evidence per check and a one-line read-back. No second model judges anything; nothing is sent.
• position_size(bankroll_usd, price, fair_value): fractional-Kelly sizing, capped, refusing negative-edge bets, returning its own assumptions.

Tools (41)

• search_markets, get_market, get_orderbook, price_history, get_positions: read-only, no keys
• overshoot_signal, premium: fresh panic-jump detection + this market's historical reversion tendency (ported from the polymarket-wc analyzer)
• dispute_risk, premium: transparent 0–100 heuristic for contested (UMA-dispute-prone) resolutions
• place_order, cancel_order, open_orders: trading, dry-run by default. price is a probability in (0,1), size is in SHARES, and the exchange enforces a $1 minimum notional on marketable orders. Trading tools carry destructiveHint annotations so clients can gate them.
• builder_stats: attribution verification + public builder leaderboard
• find_markets, compare_venues, quote_cost: cross-venue (above)
• server_info: config status, per-venue

Kalshi: kalshi_search_markets, kalshi_get_market, kalshi_get_orderbook, kalshi_get_trades, kalshi_balance, kalshi_positions, kalshi_open_orders, kalshi_place_order, kalshi_cancel_order.

Stack notes

• Official unified SDK polymarket-client (0.6.x): AsyncPublicClient for data, AsyncSecureClient.place_limit_order(..., builder_code=...) for attributed orders. The legacy py-clob-client is archived and cannot attach builder codes. Do not use it.
• MCP SDK 2.0: MCPServer from mcp.server.mcpserver (the old mcp.server.fastmcp.FastMCP import is gone in 2.x).
• Kalshi is on plain httpx + cryptography, not the official SDK: kalshi-python-sync requires Python >=3.13 and re-releases weekly in lockstep with the spec version. Auth is RSA-PSS(SHA256, salt=digest length) over str(unix_ms) + METHOD + path, where the path includes /trade-api/v2 and excludes the query string. Base URL is now external-api.kalshi.com.
• x402 (planned): the official x402 PyPI package (v2.20+) can wrap MCP tools directly (x402.mcp, payment rides in tool-call _meta), but its MCP helpers currently target mcp 1.x, so integrating means pinning mcp>=1.28,<2 or waiting for the 2.x-compatible release. Mainnet settlement needs a facilitator (Coinbase CDP: 1,000 free settlements/mo, then $0.001). Keep free tiers of both signals so registries can index the server.

Who this is for

Polymarket's public builder leaderboard shows what a single operator routing their own flow is worth. Pulled 2026-08-31 via this server's own builder_stats tool. Re-run it, the numbers move:

The instructive rows are the small ones: MagicMarkets routes $901K/week with a single active user; Jupiter $515K with one; Sharkbetting $1.15M with two. Those are bot operators routing their own flow, which is exactly who this is built for.

Roadmap

1. Live smoke test from an unblocked network: done 2026-08-23, all tools pass
2. Register builder code (polymarket.com → Settings → Builders), set fees to 0 bps at launch, export ODDSRAIL_BUILDER_CODE; first attributed order on a tiny size
3. Kalshi as venue #2: done 2026-08-23, 9 tools, verified live
4. x402 paid wrapping for the two signals once the mcp-2.x conflict clears
5. Registry listings: official MCP registry (mcp-publisher, PyPI mcp-name: marker), Smithery (needs public streamable-HTTP + a free tool for their scanner), Glama (glama.json)

Listing / distribution

• GitHub: https://github.com/hmesutozsoy/oddsrail (public, MIT)
• Glama: auto-crawls GitHub; glama.json in the repo root claims maintainership.
• PyPI: https://pypi.org/project/oddsrail/ (pip install oddsrail)
• Official MCP registry: listed as app.oddsrail/polymarket-kalshi-trading (renamed from …-arbitrage in 0.10.1; the old name is deprecated) (published 2026-08-30, status active). Re-publish after a version bump with mcp-publisher publish; keep server.json's version in step with pyproject.toml or the registry rejects it.
• Smithery: requires a public HTTPS streamable-HTTP endpoint, available once oddsrail is hosted rather than run locally over stdio.