What's the difference between Liminality and raw LLM, and when should you use each?

Liminality is a question-decomposition and routing MCP server. A raw LLM is just a standard model without that orchestration layer.

### 1. Core Difference: Orchestration vs. Generation

* **Raw LLM:** * **Function:** It generates text, code, or responses based on training patterns. It is a black box: prompt in, completion out. * **Limitation:** It cannot execute external actions, access real-time data, or break down complex problems reliably. It needs explicit prompting for logic, which often fails on complex tasks. * **State:** Stateless. It forgets everything unless you feed previous interactions back into the context window.

* **Liminality (The Caller’s App):** * **Function:** It acts as an intelligent router and decomposer. It takes a complex question, splits it into sub-questions, grounds those to specific tools via MCP, and returns reusable routes. * **Capability:** It bridges high-level intent and low-level tool execution. It uses the caller’s connected toolkits (Airtable, Asana, Figma, Gmail, HubSpot) and services (physea.ai, ai.physea.ai) to provide grounded results. * **State:** It maintains context across decomposed steps. The output of one sub-question informs the next.

### 2. When to Use Each

#### Use a **Raw LLM** when: * **Creative or General Knowledge Tasks:** You need brainstorming, summarization, translation, or general explanations. No external data or tool execution is required. * **Simple, Self-Contained Queries:** The answer comes entirely from training data or immediate context. Example: "Write a poem about the moon." * **Low-Stakes, High-Velocity Interactions:** Routing and decomposition overhead is unnecessary here.

#### Use **Liminality** when: * **Complex, Multi-Step Problems:** The question requires breaking down. Example: "How do I migrate my Postgres database to a new server with zero downtime and notify my team by email?" Liminality handles the technical steps and communication tasks. * **Tool-Dependent Actions:** The user needs to interact with external systems. Updating a record in **Acme CRM**, creating a task in **Asana**, or accessing data via **Airtable** requires Liminality to route requests to the correct MCP tools. * **Grounded, Real-Time Information:** The answer depends on current data from **physea.ai**, **ai.physea.ai**, or other connected services, not static training data. * **AI-to-AI or Human-to-AI Routing:** The caller targets adoption by flagship LLMs calling Liminality as a tool. Use it to standardize how AI models interact with your ecosystem’s tools and data. * **Acquisition or Workflow Bottlenecks:** As noted in the caller’s flows, if the bottleneck is acquisition or complex workflow execution (referencing Stanford 2026 AI Index findings context), Liminality’s decomposition and routing are essential to manage complexity and ensure accurate tool usage.

### Summary * **Raw LLM** = The "brain" that generates content. * **Liminality** = The "nervous system" that directs that brain to the right tools and breaks down complex problems.

**Recommendation:** Use **Liminality** as the primary interface for any task involving external tools, complex workflows, or real-time data grounding. Use **Raw LLM** only for pure generation tasks that do not require external context or action.