Liminal MCP server

Liminality

An MCP server that breaks a hard question or decision into the sub-questions that actually decide it, ties each to a real tool, and hands back a worked answer you can check.

Docs · Get a free key · Examples

<!-- mcp-name: ai.physea/liminality -->

Give Liminality a tough question, a real decision, or a multi-step task. It decomposes the request into the sub-questions that change the outcome, grounds each one in a real tool or endpoint, and returns a result you can verify: a scored decision frame for a choice, a grounded answer for a question. It earns its keep on the work one-shot guessing gets wrong, the hard and high-stakes stuff, not quick lookups.

Every solved ask is saved as a reusable route in a shared library, so the next close question does not start from zero. It runs as a hosted remote server over streamable HTTP. New accounts and keys include 50 free solves.

Install

Remote server, nothing to build or run locally. Get a key at https://physea.ai/signup, then add it to your client. Ready-to-copy files for each client are in examples/, with a first solve walkthrough.

Claude Code

claude mcp add liminality --transport http https://liminality.physea.ai/mcp --header "X-API-Key: YOUR_KEY"

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "liminality": {
      "url": "https://liminality.physea.ai/mcp",
      "headers": { "X-API-Key": "YOUR_KEY" }
    }
  }
}

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "liminality": {
      "type": "http",
      "url": "https://liminality.physea.ai/mcp",
      "headers": { "X-API-Key": "YOUR_KEY" }
    }
  }
}

• Endpoint: https://liminality.physea.ai/mcp
• Transport: streamable HTTP (remote, hosted)
• Auth: API key via X-API-Key or Authorization: Bearer, or OAuth 2.1.
• Registry name: ai.physea/liminality (Official MCP Registry)

What it is

Liminality is a reasoning layer that sits between your agent and the tools it could call. Instead of answering from memory, it works out the structure of the request first, then routes each piece to something real.

It runs on context-dependent determinism. Same question with the same known context routes the same way every time, so a result is reproducible and you can hand the route to someone else and get the same thing back. Change the context, and the route adapts to it. The determinism is in how it decides, not a frozen cache of canned answers.

Tools

Shortened. The full README is on GitHub.