Getting Started with qashqade's MCP Layer

Recently we introduced the qashqade MCP layer, a set of Model Context Protocol servers that let you connect qashqade to the AI tools your team already use. Instead of logging in to the qashqade interface to read a figure off a screen, your AI can ask qashqade a question in plain language, in other cases trigger a calculation, and read back validated results, all inside whatever you're already working in.

In this walkthrough, I take you through one of our MCP servers from a standing start to a finished, branded dashboard. I show what you need to install, how access to the MCP server is configured, how to phrase a request, and what the outputs look like.

‍

Below is a summary of what I cover, along with the steps to get it running yourself.

A quick word on what the MCP layer actually is

An MCP server connects an AI to an underlying system, in this case, qashqade. The server exposes one or more tools, and a skill joins those tools together to produce a meaningful end result.

The principle to keep in mind: the AI never performs the allocation maths. It asks qashqade's deterministic engine to run and reads back what the engine produced. An MCP server is simply a second door into the platform. The interface is the door a person uses, the MCP server is the door an AI uses. Every figure is still computed in the same way on every run, traceable and defensible to LPs and auditors.

Getting it running

In the demo, I use Claude Code as the AI client, paired with the skill qashqade provides. The setup breaks down into a few steps:

1. Install the qashqade skill. qashqade supplies a skill (in the demo, qashqade Insights) as a zipped file. Drop it into your AI client's skills directory and unzip it; it expands into its own folder. The next time you start the client, it automatically picks up the skill and it's ready to use.
2. Point your AI client at a working directory. When you first start Claude Code, you tell it which directory to use for temporary files. This is entirely your choice, it can be anywhere you like.
3. Get access to the MCP server. There are two routes, and which one applies depends on your organisation. Your Claude administrator can set up the environment centrally and control who has access; or, if you're working solo  (running a proof of concept) you can enable it locally for yourself. Setup instructions are provided alongside the MCP layer.
4. Start a session and sign in. Create a new session, give it a title and a task, and run it. The MCP server signs you in using the same login mechanism as the platform: it validates you as a user and returns a bearer token. That token governs what data you can see and what you're allowed to do; the AI inherits exactly your permissions, and never sees more than the person behind it. Once that's done, you're ready to ask questions.

A note for newcomers: the skill handles the heavy lifting, so you don't need to fine-tune model settings to get good results. In the demo I run on the latest available model with effort set to medium, which is more than enough.

A use case: from a simple query to a full analysis

We build up gradually to show the range, starting simple, asking the AI to retrieve all the vehicles (funds) in a development environment. The skill recognises the request, connects to the environment, and returns the available vehicles.

From there we raise the bar: find the top five best-performing funds. The AI returns the ranking and the values, quickly.

Then comes the real demonstration of what the layer can do: find the last available waterfall run, take its latest output, and produce an analysis document. Behind the scenes, the skill stitches together several of the MCP server's tools and carries out work that a person could do, but that would take far longer by hand. A couple of minutes later, it returns a finished result.

The outputs: branded dashboards, in the format you need

The headline result is the dashboard. When the waterfall analysis completes, the AI produces a dashboard that's automatically styled in qashqade branding and generates both light and dark mode versions.

The dashboard is created as HTML, which is ideal if that's how your organisation likes to work. But you're not locked into it: just ask, and the output is converted to a PDF or a Word document on the spot. In the demo, I convert a dashboard to PDF simply by asking for it; no downloads, no extra plugins, no additional services to buy.

A realistic note

As I show in this walkthrough, the experience isn't always perfectly linear, occasionally the AI doesn't locate a file first time or can't preview something, but it consistently finds a way around the issue or hands you a link to the output. These rough edges are getting smoother with every release, and from a user's perspective the task gets done and the expected output gets produced.

Want to try it?

The qashqade MCP layer works with any MCP-compatible AI, ships with ready-made skills so your tools know how to use it, and is available today for both existing and new clients. Read the full announcement here, or get in touch and we'll switch it on for you.