An unofficial, safety-focused Model Context Protocol (MCP) server for Kalshi event contracts and perpetual markets. This is a pre-release project that is not ready for trading. The server is designed for agent harnesses such as Hermes and exposes a deliberately small tool surface generated from Kalshi's published API contracts.

Canonical publisher: CopyleftDev. Planned source and container releases use github.com/copyleftdev/kalshi-kernel and ghcr.io/copyleftdev/kalshi-kernel respectively.

The independent Kernel mark follows a conservative interpretation of Kalshi's published color, contrast, clear-space, and misuse guidance. It does not use or modify Kalshi's logo. See the visual identity usage rules.

Trading APIs are broad and optimized for application developers. Agents need a narrower boundary with explicit intent, strict schemas, predictable errors, and strong separation between simulation and real-money execution. Kalshi Kernel provides that boundary while continuously testing its generated clients against the latest upstream contracts.

The design priorities are:

• paper mode by default;
• immutable execution mode for the lifetime of a process;
• explicit acknowledgement and credentials before live mode can arm;
• fixed-point strings for prices and quantities;
• strict, generated tool schemas with safety annotations;
• fail-closed behavior when a capability is unavailable;
• traceable alignment with exact OpenAPI operations and AsyncAPI channels; and
• repeatable MCP protocol, race, fuzz, and upstream-drift testing.

The table describes the current main branch, not the published v0.1.1 tag. The additions are unreleased and the version metadata has not yet advanced.

Treat the current branch as development software, not as a published or audited trading integration. Some design and publication documents still describe the v0.1.1 scaffold; compare them with the implementation before relying on a claim. See ARCHITECTURE.md, THREAT_MODEL.md, and the release gates in PUBLICATION.md.

Connected MCP clients also receive server-level instructions telling them to call kernel_status first, distinguish paper from live mode, and never report an action as successful unless its structured response contains ok: true. Until its readiness reporting is updated, confirm live arm state from the arm_live_trading response rather than from kernel_status.

Prerequisites:

• Go 1.25.13 or newer;
• Node.js 22 only for the official MCP conformance runner; and
• a supported MCP client.

make test
make build
KALSHI_KERNEL_MODE=paper ./bin/kalshi-kernel

Paper mode is the default when KALSHI_KERNEL_MODE is unset. It intentionally discards live credential configuration.

Add a local stdio server to ~/.hermes/config.yaml:

mcp_servers:
  kalshi-kernel:
    command: "/absolute/path/to/kalshi-kernel/bin/kalshi-kernel"
    env:
      KALSHI_KERNEL_MODE: "paper"

Restart Hermes and call kernel_status before any other tool.

After building the binary:

claude mcp add --scope user --transport stdio kalshi-kernel \
  --env KALSHI_KERNEL_MODE=paper -- /absolute/path/to/bin/kalshi-kernel

Then run claude mcp get kalshi-kernel or open /mcp to confirm the connection. This local setup is separate from Anthropic's public Connectors Directory, which requires a deployed remote server.

Use the standard MCP configuration shape:

{
  "mcpServers": {
    "kalshi-kernel": {
      "command": "/absolute/path/to/bin/kalshi-kernel",
      "env": {
        "KALSHI_KERNEL_MODE": "paper"
      }
    }
  }
}

KALSHI_KERNEL_MODE=paper ./bin/kalshi-kernel

Paper mode uses an in-memory simulated ledger and never submits an order to Kalshi. It starts with $100.00 by default; set KALSHI_PAPER_CASH_DOLLARS before startup to choose another balance. The ledger, positions, idempotency records, and fill journal reset when the process exits.

place_order fetches a fresh public order-book snapshot and supports only immediate, all-or-nothing marketable fills. The requested price must exactly equal the current touch and the displayed size must cover the full quantity. The simulator uses fixed-point arithmetic, applies the published fee formula, records a hash of the book used, and treats client_order_id as an idempotency key. Because paper orders never rest, amend_order and cancel_order return typed failures.

Every paper response is labeled simulated: true. Simulated fills do not predict or guarantee live fills, liquidity, latency, slippage, fees, or profitability.

Live mode is deliberately difficult to enable:

KALSHI_KERNEL_MODE=live
KALSHI_API_KEY_ID=...
KALSHI_PRIVATE_KEY_PATH=/absolute/path/to/private-key.pem
KALSHI_LIVE_TRADING_ACK=I_UNDERSTAND_THIS_TRADES_REAL_MONEY

# Optional startup-only overrides; these are the defaults.
KALSHI_MAX_ORDER_NOTIONAL_DOLLARS=25.00
KALSHI_MAX_DAILY_NOTIONAL_DOLLARS=100.00
KALSHI_MAX_DAILY_ORDERS=200

Startup configuration selects immutable live mode and enables authenticated event-contract portfolio reads and cancellation. It does not authorize place/amend calls. Those calls require a second, process-local MCP action:

{
  "acknowledgement": "I_UNDERSTAND_THIS_TRADES_REAL_MONEY",
  "arm": true
}

Pass that object to arm_live_trading; pass the same object with arm: false to disarm. Arming is never persisted. place_order also requires a stable, caller-provided client_order_id. Place/amend requests are rejected before submission when they exceed the startup-only per-order, UTC-day notional, or UTC-day order-count limits. These counters are process-local and reset when the process restarts.

Live account reads and writes currently support event contracts only. Perpetual-account operations and non-primary subaccounts are unavailable; omit subaccount or pass 0. Network failures after a write produce reconciled or explicitly indeterminate outcomes rather than blind retries.

Never commit credentials, private keys, or .env files. Use short-lived credentials where available, a secrets manager in production, and an account with only the privileges and capital required.

The curated agent interface lives in specs/mcp-tools.yaml. It maps each MCP tool to exact OpenAPI operation IDs and AsyncAPI channels. Administrative and account-management endpoints are not automatically exposed merely because they exist upstream.

Four authoritative contracts are fetched from https://docs.kalshi.com:

• openapi.yaml — event-contract REST API;
• asyncapi.yaml — event-contract WebSocket API;
• perps_openapi.yaml — perpetuals REST API; and
• perps_asyncapi.yaml — perpetuals WebSocket API.

The fetcher pins HTTPS and the source hostname, limits response sizes, validates the expected dialect and non-empty surface, and records hashes and HTTP provenance. The upstream contract bodies are ignored by Git and fetched into a local cache as needed. Upstream-derived REST clients are also generated locally and excluded from source releases. Only the provenance lock, curated MCP overlay, and generated MCP tool boundary are versioned. Regeneration and testing happen in a temporary repository copy before a tested lock and MCP boundary can be promoted.

make upstream-test      # test latest contracts without requiring snapshot parity
make upstream-check     # require parity and run MCP conformance
make upstream-promote   # promote only the exact artifacts that passed

Every CI run performs the strict upstream check; CI also runs daily so drift is detected even when the repository is idle.

make test              # unit, integration, and contract alignment tests
make test-alignment    # OpenAPI, AsyncAPI, and MCP overlay alignment
make check-generated   # deterministic generated-code check
make test-race         # concurrent calls under Go's race detector
make test-fuzz         # execution-mode parser safety fuzzing
make conformance       # pinned official MCP protocol scenarios
make release-test      # complete local release gate
make public-check      # vulnerability, metadata, and public-readiness gate
make upstream-check    # latest Kalshi contracts plus conformance

The suite checks transport security, authentication contracts, mutation security, operation-ID uniqueness, fixed-point types, curated surface area, required-field parity, strict JSON Schemas, tool titles and safety hints, fail-closed behavior, concurrency, and Streamable HTTP negotiation.

The repository contains publication metadata for the official MCP Registry and submission dossiers for the OpenAI Plugins Directory and Anthropic directories. Metadata is preparation—not evidence of approval, endorsement, or publication.

Anthropic's current connector review criteria do not accept connectors that transfer money or other financial assets. Any Claude Connectors Directory build of this project must therefore be a distinct read-only and/or paper-only artifact with live order tools omitted, subject to Anthropic's review.

See PUBLICATION.md for the exact channel matrix, current blockers, release steps, listing copy, and review test cases. No external registry submission is performed automatically.

Read CONTRIBUTING.md before opening a pull request. Report vulnerabilities privately according to SECURITY.md; do not open a public issue for a suspected credential leak or trading-safety vulnerability. General usage help belongs in SUPPORT.md.

This software is provided for development and research. It does not provide investment, financial, legal, tax, compliance, or trading advice. Event contracts and leveraged products can result in rapid and substantial loss, including loss of the entire amount committed. You are responsible for account eligibility, jurisdictional restrictions, exchange rules, regulatory requirements, taxes, strategy, orders, and losses.

AI systems and software can misunderstand intent, produce incorrect parameters, repeat requests, or behave unexpectedly. Human review, exchange-side risk limits, least-privilege credentials, monitoring, and an independent emergency stop are required before any live deployment.

Read the full DISCLAIMER.md, PRIVACY.md, and SECURITY.md. The project is licensed under Apache License 2.0; third-party material is excluded as described in THIRD_PARTY_NOTICES.md.