mcpdirectory.dev
9,129 servers · updated 23 Sept

Liminality MCP server

community Active last commit 15 d
{
  "mcpServers": {
    "physea-liminality": {
      "type": "http",
      "url": "https://liminality.physea.ai/mcp"
    }
  }
}
file .mcp.json (in your project root)version 1.0.5

What it does

Liminality is an MCP server that decomposes complex questions and decisions into sub-questions, grounds each to real tools, and returns a verifiable answer or scored decision frame. It is for developers building agents that need to handle high-stakes, multi-step reasoning tasks with reproducible results.

  • Break down a complex business decision into testable sub-questions, each routed to relevant data sources or tools, and receive a scored frame showing how each factor affects the outcome.
  • Get grounded answers to research questions by having the system determine which tools to call, execute them, and synthesize results you can verify and trace.
  • Solve multi-step problems where the path to the answer is not obvious, with deterministic routing so the same question with the same context produces the same result every time.

Tools this server exposes

ToolDoes
solveDecomposes a question or decision into sub-questions, routes each to real tools, and returns a worked answer or scored decision frame.
researchRuns a multi-source pass to gather real information for a question.
ask_formReturns a short form when the answer depends on user-specific details.
apply_formFolds form responses into a refined result.
get_my_contextRetrieves what is known about the user and connected tools.
register_assetRegisters user material for use in routing and answers.
set_preferenceSets how the user prefers results to be formatted or delivered.
report_feedbackProvides feedback on a result to improve future routing.
report_outcomeReports how a result performed in practice.
composio_connectConnects a tool via Composio so actions can run against it.

Before you enable it

An MCP server runs with whatever access you give it. Start with a development account rather than production, and check which tools the server exposes before you let an agent call them unattended.