The desk as tools

SKEW ships a Model Context Protocol server. MCP is the open standard that lets an AI client call a program’s functions as typed tools: you register the server once, and Claude can then scan for mispriced volatility, read a stress grid, inspect the risk tier or replay any decision — by asking, in plain language. Reads are always available. The two mutating tools are not even registered unless explicitly enabled, and both re-run the full gate chain and require an explicit confirmation argument. The model gets a desk, never a blank cheque.

Setup — three commands, copy-paste

1. 1 · Clone and install

   git clone https://github.com/SooryaCodes/skew.git ~/skew
   cd ~/skew/backend
   python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

2. 2 · Add your Alpaca paper keys

   cp ~/skew/.env.example ~/skew/.env
   # edit ~/skew/.env: ALPACA_API_KEY, ALPACA_API_SECRET (paper keys, free at alpaca.markets)
   ~/skew/backend/.venv/bin/python -m skew.cli account   # confirms the account + options level

   The server refuses to start against anything but the paper endpoint, so there is nothing dangerous to misconfigure.

3. 3 · Register it with your client

   The exact blocks for Claude Code and Claude Desktop are below — with ~/skew paths that match step 1, so they work as pasted.

Read tools — always on

• scan_volatility(symbols?)

  Volatility state for the universe — IV, RV, VRP, regime per name.

• propose_structures(symbol)

  Defined-risk candidates for one name, sized to the risk budget.

• stress_test(candidate_id)

  The 84-scenario grid for a candidate: price × IV × time, breaches marked.

• risk_status()

  Tier, per-trade and portfolio budgets, headroom, drawdown.

• positions()

  The open book, marked to market, with exit conditions.

• audit_log(limit?, action?)

  Every decision with its reason — refusals as prominent as fills.

• desk_status()

  Armed state, market clock, account (paper), last cycle.

Write tools — off unless enabled, confirm-required

• execute(candidate_id, confirm)

  Submit one gated candidate as an atomic multi-leg order. Requires confirm=true and runs the full gate chain again first.

• close(position_id, confirm)

  Close one open structure as a single order. Same confirmation contract.

Claude Code — one line

claude mcp add skew -- ~/skew/backend/.venv/bin/python -m skew.mcp_server

Then /mcp inside a session to confirm it connected.

Claude Desktop

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "skew": {
      "command": "/Users/YOUR_USERNAME/skew/backend/.venv/bin/python",
      "args": ["-m", "skew.mcp_server"],
      "cwd": "/Users/YOUR_USERNAME/skew/backend"
    }
  }
}

Claude Desktop needs absolute paths (it does not expand ~) — replace YOUR_USERNAME, and keep the command pointing at the virtualenv’s Python: a bare python will not have alpaca-py installed. Restart Claude Desktop and SKEW appears in the tools menu.

Try this — under a minute

1. “What’s the desk’s status?” — armed state, market clock, tier, paper-only guarantee.
2. “Scan volatility and tell me which name has the widest gap between implied and realized.” — the live VRP table, regime per name.
3. “Propose structures for that name and stress-test the first one.” — real candidates sized to the budget, then the 84-cell grid with any breaches marked.
4. “Show the last five decisions and why.” — the audit log, refusals included, each traceable.