dcisionai-mcp-client

Local stdio MCP server that forwards every JSON-RPC call to the DcisionAI hosted streamable-HTTP MCP endpoint at https://platform.dcisionai.com/mcp (or wherever DCISIONAI_MCP_URL points).

Why this exists

The DcisionAI optimization pipeline (Domain Agent → Consultant Agent → solver → explanation, with the Solver Intelligence layer) runs as a multi-service deployment on Railway. Some MCP hosts (older Claude Desktop / Cursor builds, sandboxed environments) only accept stdio MCP servers launched via command + args. This package is the official stdio↔HTTP bridge for those hosts.

For newer MCP hosts that speak streamable-HTTP natively (Claude Desktop ≥1.0, Cursor ≥0.42, Claude Code ≥0.5), use the direct URL config instead — there's no advantage to running this proxy.

Install

# One-shot launch (recommended for MCP host configs)
uvx --from dcisionai-mcp-client@latest dcisionai-mcp-client

# Or pin a version
uvx --from dcisionai-mcp-client==3.4.0 dcisionai-mcp-client

# Or install globally
pipx install dcisionai-mcp-client
dcisionai-mcp-client

MCP host configuration

Claude Desktop / Cursor / Claude Code

Add to your mcpServers config:

{
  "mcpServers": {
    "dcisionai-optimization": {
      "command": "uvx",
      "args": [
        "--from",
        "dcisionai-mcp-client@latest",
        "dcisionai-mcp-client"
      ],
      "env": {
        "DCISIONAI_MCP_URL": "https://platform.dcisionai.com/mcp"
      },
      "disabled": false,
      "autoApprove": [
        "preflight_optimization",
        "submit_optimization_job",
        "get_job_status",
        "get_job_result",
        "get_job_raw",
        "get_job_thinking",
        "get_hitl_questions",
        "respond_to_hitl",
        "run_optimization_headless"
      ]
    }
  }
}

env.DCISIONAI_MCP_URL is optional — omit to default to production.

URL by environment

Environment variables

What it does NOT do

This proxy is a transport adapter, not a local solver. The optimization work — Domain Agent, codegen, Pyomo/HiGHS solve, Solver Intelligence, explanation generation — all runs on the hosted endpoint. No Anthropic / Pinecone / Neo4j / Voyage / Supabase keys are needed locally. That's the whole point.

If you need a self-contained server (the pipeline running entirely on your machine, BYO credentials, no network dependency on DcisionAI infrastructure), look at dcisionai-mcp-server@3.3.0 (separate heavyweight package). The two coexist; the -client package is the recommended path for the vast majority of integrations.

Tool surface

The 9 tools the remote endpoint exposes (you'll see these in your MCP host's tool panel after a successful handshake):

Tool surface is authoritative on the server (mcp/public/server.py); the proxy doesn't filter or transform — whatever the remote advertises is what your host sees.

Architecture

┌─────────────────┐   stdio JSON-RPC   ┌─────────────────────────┐
│   MCP host      │ ◄────────────────► │   dcisionai-mcp-client  │
│ (Claude Desktop,│                    │   (this package)        │
│  Cursor, etc.)  │                    └────────────┬────────────┘
└─────────────────┘                                 │ HTTPS POST + SSE
                                                    │ /mcp
                                                    ▼
                                        ┌─────────────────────────┐
                                        │   DcisionAI platform    │
                                        │   (Railway / Cloud)     │
                                        │  Domain → Consultant →  │
                                        │   solver → explain      │
                                        └─────────────────────────┘

Each stdin line is a JSON-RPC message. The proxy POSTs it, captures Mcp-Session-Id from the initialize response, forwards it on subsequent requests, parses the SSE response body, writes the JSON-RPC payload to stdout.

License

Proprietary. © DcisionAI.