## Why This is Important

The Harness can accept complex Input. The Report it generates is also complex, and contains suggestions for how results can be improved. This is an iterative process that is much more powerful than trying to get perfect results with a single shot.

In the AI era the natural approach is to use an LLM both to prepare the Input and interpret the Report in successive iterations. This can best be accomplished by using the Harness as an AI tool, using well-established methods as set out below.

The web version of the Harness is maintained principally in order to: a. showcase how the Harness works; b. contain information necessary to understand the Harness; c. provide the means by which Accounts and billing can be managed. 

In summary it is strongly recommended that actual use of the Harness is as an AI tool. 

## Primary Chat

The primary chat is between the user and their LLM, which will use their context to compose the input to the Harness, and use Harness results to provide answers and further refine the input. This iterative process, combining context knowledge with specialist knowledge provided by the Harness, provides high value.

## One-Shot, No Session

Each call of the Harness by the user's LLM is independent. The user's LLM controls the memory of previous calls. This ensures an appropriate separation of concerns, and reinforces least agency.

## Billing

Each call is charged the same flat rate as a report generated through the web form, immediately on completion. There is no separate pricing for integrations.

## MCP Integration

### Overview

The recommended integration path is via Model Context Protocol (MCP).  MCP is supported by Claude, OpenAI, Gemini, and most desktop/IDE MCP clients (Cursor, Windsurf, Cline, and others), so may be regarded as a dominant open standard.

Connecting the Harness to AI is a one-off task for whoever manages your firm's use of AI tools. Once the Harness is connected, your AI tool will decide whether and how to use it depending on user input. 

### Installation

For installation, configuration, and guidance the quickest way is to enter the following prompt into your AI chat window:  **Please go to https://github.com/99blakeD99/the-metis-fca-handbook-ai-harness-mcp-files/blob/main/README.md#quick-start and follow the instructions for installing the MCP server.**

To use the local (stdio) server instead (for clients that cannot use the remote endpoint or users who prefer not to), add: **Do not use the remote endpoint; use the local (stdio) server described under "Alternative: run the server locally (stdio)".**

The link also works if you want to do the work manually.

## Non-MCP Integrations

We are currently evaluating additional integration methods (HTTP APIs, webhooks, plugins, etc.). We will document and support these once they have passed the same testing and maturity process as the MCP integration.

## MCP and A2A Protocols

MCP and A2A are the principal protocols for agentic interaction, but at different layers. A2A coordinates communication between AI agents; MCP connects an agent to external tools and data sources. The Harness is such a tool, accessed via MCP. It does not participate in A2A orchestration, and cannot be drawn into it. This preserves separation of concerns and keeps the Harness's agency to the minimum needed to do its job.