OptionsAhoy MCP Server

Independently verified by third parties. Glama: third-party MCP-directory quality score (tool docs, behavior, completeness). · npm: published with build provenance, a signed SLSA attestation that this package was built from this repo by GitHub Actions (verify with npm audit signatures). · MCPSafe: independent 5-model-consensus security scan (AIVSS), Grade A with zero findings.

Validated against trusted sources (checks we run ourselves, against references we do not control, and that you can reproduce). Computation: every 2026 federal tax constant matches its IRS Rev. Proc. 2025-32 / Internal Revenue Code value, and 14 worked federal cases (ordinary income, long-term capital gains, and the Alternative Minimum Tax including the incentive stock option bargain element) reproduce to the cent against the independently-maintained PSL Tax-Calculator, a tax model we did not write. State income tax is cross-checked the same way: 16 cases across California, New York, New Jersey, Pennsylvania, and Massachusetts reproduce to the cent against OpenTaxSolver, an independent state tax engine we also did not write. The headline answer is recomputed live in your browser.

Tested and hardened. Input safety: requests are validated against the published schema; bad inputs return a clear 400 with the offending field named, never a crash or a wrong number, and the live API is re-checked by a robustness suite after every deploy. · Test suite: the calculation engine is covered by more than a thousand automated tests across the federal and 50-state tax logic, AMT credit recovery, and option pricing; a failing test blocks the release.

Live usage: MCP calls over the last 30 days, served straight from the server's own telemetry (/api/v1/stats, aggregate counts only, no PII).

Deterministic equity-compensation tax math that any Model Context Protocol (MCP) client can call: incentive stock option (ISO) exercise schedules under the alternative minimum tax (AMT), non-qualified stock option (NSO) and restricted stock unit (RSU) decisions, qualified small business stock (QSBS) qualification, single-stock concentration, protective-put hedging, and equity-funding goals. Full federal tax code plus all 50 states and DC, 2026 brackets. Built by AlphaLatitude Inc., the company behind OptionsAhoy.

Why not just ask the model? We benchmarked five frontier large language models (LLMs), 3 runs each, 15 trials total, on the same multi-year ISO exercise problem. Every trial overshot the true after-tax outcome, by 2x to 20x. Multi-year scheduling has a search space larger than an LLM can reason through in-context; these tools return the verifiable answer instead. Live benchmark, updated for the latest models: optionsahoy.com/benchmark. Raw responses and scoring: llm-iso-benchmark. Full write-up: But can it do taxes though?

Install in one line

The hosted endpoint is https://optionsahoy.com/mcp (HTTP, no auth, no account). Quickest paths:

Full install matrix (Gemini CLI extension, config-file JSON, REST API, Google Cloud Agent Registry): optionsahoy.com/for-agents.

Related MCP server: irs-taxpayer-mcp

The seven tools

The optimizers return the best schedule across the candidate space; the calculators return exact deterministic results. Deterministic computation, not a language-model guess, validated against brute-force ground truth on tractable cases (see the proof). Coverage spans the full federal tax code (ordinary brackets, long-term capital gains, AMT with credit recovery, FICA, NIIT) plus all 50 states and DC (state ordinary brackets, LTCG treatment, state AMT for CA, CO, CT, MN). Same engine as the in-browser calculators at optionsahoy.com/tools; the API response is byte-identical to clicking through the tool.

Use it in your agent framework (Python)

If you build agents in Python rather than calling the MCP endpoint directly, OptionsAhoy ships installable tool packages for the major agent frameworks. Each one wraps the same calculators behind the framework's native tool interface. All are published on PyPI and all are keyless: no OptionsAhoy account, no API key.

The three framework adapters pull in the keyless optionsahoy client automatically. There is also an OpenBB Workspace agent (a FastAPI application built on the OptionsAhoy client) for use inside OpenBB Workspace. Source and runnable examples for all of the above live under integrations/python.

More ways to build

However your agent is built, there is a drop-in piece. All are public and keyless.

Try it without installing

The live widget on optionsahoy.com/for-agents calls this same endpoint from your browser. No client, no config.

Prefer a chat interface? The same calculators answer plain-language questions at poe.com/OptionsAhoy.

Or watch a real session:

Real Claude Code session, unedited. A multi-stack META question (10K ISOs + 6K vested RSUs + 2K fresh RSUs + $400K house in 2027) fires 4 OptionsAhoy MCP tools in parallel: concentration risk, equity funding plan, AMT/ISO optimization, protective put pricing. Claude synthesizes the outputs into one plan that overrides each tool's standalone pick because the user is 86% concentrated in META. 2:13. Click the poster to play it on optionsahoy.com.

Endpoints and discovery

Live MCP endpoint: https://optionsahoy.com/mcp Live REST API: https://optionsahoy.com/api/v1 OpenAPI 3.1 spec: /openapi.json Discovery manifests: /.well-known/mcp.json · /.well-known/openapi.json Agent integration docs: optionsahoy.com/for-agents

MCP resources (topical briefings)

Seven markdown resources under resources/list give an LLM enough grounding to discuss the topic before picking a tool. Most map 1:1 with a cornerstone article on optionsahoy.com/learn and the matching calculator; the equity-funding briefing maps to its calculator.

MCP prompts (workflow scaffolds)

Seven prompts under prompts/list scaffold typical user questions and route to the right tool. In Claude Desktop they appear as named slash-commands; in any MCP client, prompts/get { name, arguments } returns a fully-templated user message.

Install details

Claude Desktop extension (one-click)

The optionsahoy.mcpb bundle installs by double-click (or drag onto Claude Desktop → Settings → Extensions), with no terminal or config-file editing, using Claude Desktop's built-in Node.js runtime.

To build the bundle from source:

npm install && npm run build:mcpb

Smithery CLI (19 clients, one command)

npx @smithery/cli install alphalatitude/optionsahoy --client claude

Swap claude for any client Smithery supports: claude-code, cursor, vscode, gemini-cli, codex, windsurf, cline, goose, opencode, and 10 more. Listing: smithery.ai/servers/alphalatitude/optionsahoy.

Gemini CLI extension

gemini extensions install https://github.com/AlvisoOculus/optionsahoy-mcp

This repo doubles as a Gemini CLI extension: gemini-extension.json wires the hosted MCP endpoint and GEMINI.md provides usage context to the model.

Local stdio (npm)

For clients that only support local stdio servers (Claude Desktop without mcp-remote, some IDE integrations):

npx -y optionsahoy-mcp

Or add to a Claude Desktop / Cline / Goose config file:

{
  "mcpServers": {
    "optionsahoy": {
      "command": "npx",
      "args": ["-y", "optionsahoy-mcp"]
    }
  }
}

The local server returns byte-identical responses to the hosted endpoint at https://optionsahoy.com/mcp. Source for both lives in functions/_lib/mcp-tools.ts; the stdio entry point is src/stdio-server.ts.

Use the REST API directly

# List endpoints
curl https://optionsahoy.com/api/v1

# Run an optimization
curl -X POST https://optionsahoy.com/api/v1/amt-iso \
  -H "content-type: application/json" \
  -d @input.json

Request body shapes are documented in public/openapi.json.

Repository layout

functions/         Cloudflare Pages Functions (MCP server + REST API endpoints)
  mcp.ts           HTTP MCP server
  api/v1/*.ts      Seven tool endpoints + stats + GET /api/v1 discovery
  _lib/*.ts        Shared helpers, calc-input parsers, MCP tool descriptors
lib/               Optimizer + tax-code logic
  calc/            Per-tool optimizer functions (computeAmtIso, etc.)
  tax/             Federal + 50-state + DC bracket data, AMT, FICA, NIIT
  markets/         Sector statistics
  options/         Black-Scholes, risk-free rates
  data/            Type definitions for option-chain data
public/            Static assets: OpenAPI spec, llms.txt, discovery manifests
tests/             Vitest suites (an extensive test suite including byte-identity assertions)

Run tests

npm install
npm test         # an extensive test suite, ~3s on a laptop
npm run typecheck

Registry listings

• Official MCP Registry — io.github.AlvisoOculus/optionsahoy-mcp v1.0.0, status active

• Smithery — alphalatitude/optionsahoy (plus the equity-plan skill)

• Gemini CLI extensions gallery — @AlvisoOculus/optionsahoy-mcp

• add-mcp curated registry

• PulseMCP (cascades from Official Registry)

• Continue.dev hub — block YAML lives at .continue/mcpServers/optionsahoy.yaml

Use from Google Cloud (Gemini agents)

Google Cloud Agent Registry lets each GCP project register external MCP servers for use by Gemini agents. Registration is per-project (no central submission). Two paths:

# Path A: let the Agent Registry introspect our MCP endpoint
gcloud alpha agent-registry mcp-servers register \
  --uri=https://optionsahoy.com/mcp \
  --display-name="OptionsAhoy" \
  --location=us-central1 \
  --import-tools

# Path B: pass our published toolspec.json directly (faster, no introspection)
gcloud alpha agent-registry mcp-servers register \
  --uri=https://optionsahoy.com/mcp \
  --display-name="OptionsAhoy" \
  --location=us-central1 \
  --tool-spec=<(curl -sSL https://optionsahoy.com/toolspec.json)

The toolspec.json mirrors the MCP tools/list response with readOnlyHint and idempotentHint annotations on all seven tools (all are pure deterministic calculators with no side effects). To regenerate after a tool-shape change:

curl -sS -X POST https://optionsahoy.com/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}' \
  | jq -c '{tools: [.result.tools[] | . + {annotations: {readOnlyHint:true, idempotentHint:true, destructiveHint:false, openWorldHint:false}}]}' \
  > public/toolspec.json

Troubleshooting

Connection refused / 404 from the MCP endpoint https://optionsahoy.com/mcp requires POST with content-type: application/json and a JSON-RPC body. A GET returns a JSON server description; any other verb returns 405. Verify with:

curl -X POST https://optionsahoy.com/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","method":"initialize","id":1,"params":{}}'

Tool calls fail with Error: ... text in the response The MCP server returns isError: true with a human-readable message when input validation fails. Most common: a required field missing, or a number passed as a string. Check the input against the inputSchema returned by tools/list, or against /openapi.json.

Tool not appearing in Claude.ai or Claude Desktop

• Confirm the connector URL is exactly https://optionsahoy.com/mcp (no trailing slash, no /v1).

• In Claude Desktop, restart the app after editing claude_desktop_config.json.

• In Claude.ai, the connector toggle is per-chat: enable it in the attachments menu.

• Check the live tools/list response (seven tools expected): curl -X POST https://optionsahoy.com/mcp -H 'content-type: application/json' -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'

CORS errors from a browser-based client The server returns access-control-allow-origin: * on all responses including preflight, and accepts the standard MCP headers (content-type, mcp-session-id, mcp-protocol-version). If a browser still blocks, the client is likely sending a non-allowed header — verify the request headers against the access-control-allow-headers response.

Resource / prompt not found Resource URIs and prompt names are case-sensitive. Pull the canonical list with resources/list and prompts/list rather than hand-typing.

Stale tax-year math The tax engine ships with 2026 inflation-adjusted brackets, OBBBA 2026 QSBS rules, and current state-conformity tables. If results look off for a multi-year horizon, verify the input grantDate, acquisitionDate, or saleDate falls in the year you expect — the engine resolves brackets per tax year.

Reporting a calculation bug or unexpected output Email andrew@alphalatitude.com with: the exact JSON-RPC request body, the response, the expected value, and (if known) the IRS publication or state statute the expected value derives from.

Privacy Policy

Full policy: optionsahoy.com/privacy.

In short: no account is required and no personally identifiable information is stored — no name, email, IP address, or login. Tool inputs and outputs are retained briefly (about seven days) for debugging and product improvement, alongside aggregate usage metadata (tool, timestamp, coarse location, client type) used to understand usage and detect abuse. The local stdio server and the Claude Desktop extension compute everything on your machine; the only network request is an option-chain lookup (ticker symbol only) for protective_put_price.

License

MIT. See LICENSE. The deployed service at https://optionsahoy.com/mcp and https://optionsahoy.com/api/v1 is free during beta under terms.

Contact

For partnerships, early API access, MCP integration support: andrew@alphalatitude.com