@shiftyasan/mcp-server

Model Context Protocol (MCP) server for the Shiftyasan public solver API. Lets AI agents (Claude Desktop, Cursor, Continue, etc.) call the shift-scheduling solver from any MCP-aware host.

Status: open beta — validate_shift_input, solve_shift, and get_solve_job (async polling for large problems). Usage queries (get_usage) coming next.

Most people no longer need this package

Shiftyasan now runs a remote MCP server, which is the recommended way to connect:

https://api.shiftyasan.com/mcp

Connecting from Claude.ai or ChatGPT opens a Shiftyasan sign-in page in your browser; approve it once and you are done. Revoke a connection any time from the dashboard under "Connected apps".

This package is still supported and existing setups keep working. Reach for it when you want the server to run locally — for example on a host that cannot reach the internet directly, or when you prefer to hold your own API key.

See the quickstart for the remote setup.

What is Shiftyasan?

Shiftyasan is a SaaS that auto-generates work shifts for businesses (retail, hospitality, healthcare, etc.) using a constraint-optimization engine. The same solver is exposed as a public HTTP API at https://api.shiftyasan.com/v1/public/*. This MCP server is a thin client wrapper that turns the API into discoverable tools for AI agents.

Developer documentation — quickstart, the full list of supported constraints, and pricing — lives at https://platform.shiftyasan.com (Japanese and English).

Tools

The full SolveRequest / SolveResponse schema lives in the OpenAPI spec at GET <base URL>/v1/public/openapi.json (base URL = SHIFTYASAN_BASE_URL, default https://api.shiftyasan.com).

Prompts

Both arguments (period, context) are optional, so it can be invoked bare. In Claude Code it shows up as /mcp__shiftyasan__plan_shift; other hosts surface prompts in their own UI.

Use it when you want the agent to interview you for the inputs rather than guessing them, or when a previous attempt produced a roster you could not check at a glance. The prompt tells the agent to render one row per person with the per-person totals on the right, which is what a shift manager reads first.

The server also sends workflow guidance in its initialize response (call validate_shift_input before solve_shift, poll get_solve_job on 202, how to read shortage markers and the infeasibility diagnosis). Hosts may add this to the model's system prompt; Claude Code does, Claude Desktop currently does not, which is why the same ordering rules also live in the tool descriptions.

Constraints the schema accepts but the solver does not apply yet

A few fields are accepted by the API and then dropped during translation, because the solver has no equivalent. validate_shift_input and solve_shift print an explicit warning listing exactly what was dropped, so an agent never reports them as honored.

Everything else in the schema is applied by the solver.

Install

Running locally requires Node.js 20 or newer. If you would rather not manage that, use the remote server instead.

Use directly via npx (no global install needed):

npx -y @shiftyasan/mcp-server

Also listed in the MCP Registry as com.shiftyasan/mcp-server.

Or install globally:

npm install -g @shiftyasan/mcp-server
shiftyasan-mcp-server

Setup

1. Get an API key

Issue an sk_live_... key in the self-service dashboard: https://platform.shiftyasan.com/dashboard/ (your existing Shiftyasan account works; sign-up is also available). If anything goes wrong, email info@shiftyasan.com (Japanese or English).

Treat the token like a password: it grants access to your solver quota.

Beta note: you may receive a base URL different from the default — set it via SHIFTYASAN_BASE_URL (see below).

2. Configure your MCP host

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "shiftyasan": {
      "command": "npx",
      "args": ["-y", "@shiftyasan/mcp-server"],
      "env": {
        "SHIFTYASAN_API_KEY": "sk_live_..."
      }
    }
  }
}

Restart Claude Desktop. The shiftyasan server appears in the tools menu and exposes validate_shift_input, solve_shift, and get_solve_job.

Claude Code

claude mcp add shiftyasan -e SHIFTYASAN_API_KEY=sk_live_... -- npx -y @shiftyasan/mcp-server

(Add -e SHIFTYASAN_BASE_URL=... if you were given a beta base URL.)

Cursor / Continue / other MCP hosts

Add to your MCP server config in the same shape (command + args + env). Refer to your host's docs for the exact config file path.

Using from ChatGPT

ChatGPT cannot spawn local stdio MCP servers like this package, so use the HTTP API directly through a Custom GPT with Actions (the OpenAPI spec is designed to be Actions-compatible — fully inlined schemas, no external refs):

1. ChatGPT → My GPTs → Create a GPT → Configure → Actions → Import from URL, and enter <base URL>/v1/public/openapi.json (base URL = the one you received with your API key; default https://api.shiftyasan.com).

2. If you were given a beta base URL, also edit the imported schema's servers[0].url to <base URL>/v1/public.

3. Authentication → API Key → Auth Type Bearer → paste your sk_live_... token.

4. Suggested GPT instructions:

   You can build optimized work shift schedules with the Shiftyasan actions. Always call the validate action first (free), then solve. Large problems return 202 with a job_id — poll the job endpoint every ~15 seconds until the status is "completed" or "failed".

Notes: creating custom GPTs requires a paid ChatGPT plan, and the API key is stored inside the GPT — do not share that GPT publicly.

ChatGPT's MCP connectors only support hosted MCP endpoints (Streamable HTTP), not local stdio processes. A hosted MCP endpoint is on our roadmap; until then, use Actions as above.

Environment variables