Comments

Nixc0rdant, I'm happy to share anything useful.  Just to be clear, I am not a programmer so I do everything with AI.  At the moment everything is with claude but I will look later at local models.  I also  follow 'knowledge packs' approach.  My initial plan was to have bfexplorer running on various staretegies and a URL to open markets and place trades.  This is working well but everything is still at testing as a few minor issues so now I'm using the AI to remove those and to increase efficiency.

Nixc0rdant, which programming language are you using for your Betfair projects?

While different languages might be used, I find that statically typed programming languages like F# and .NET languages such as C# or Visual Basic do not require loading an entire source code codebase.

For instance, with F#, we have MCP servers that can explore any type of referenced library. This approach prevents hallucination and avoids having to process millions of tokens just to browse the entire codebase.

I did not write my own prompts; I had the LLM generate them, which you can see in experiments on my GitHub.

I use DeepSeek models, which offer a 1M context window, enough for an entire session within my Betfair strategy project.

In fact, when debugging with an LLM against real-time data, the amount of that data is often larger than my entire strategy project's codebase. After testing over three days, I was genuinely surprised that given all the code DS generated into the output context, it remembered closed market data when asked to calculate P&L from our tests.

These closed markets were not within the bfexplorer application itself; rather, Reasonix/DeepSeek LLM retained this information in its memory context.

That sounds like an fun build! I’ve been deep in Claude Co‑Work and code workspace after running into similar limitations with long‑term memory and context drift. What’s helped me a lot is creating modular Markdown 'knowledge packs' that define the agent’s role, constraints, and behaviour. Cowork handles these really well when you have it reload them each session

Keen to collaborate with you on this custom agentic system and compare notes! If you’re open to it, I’d love to take a look at your project folder and README.md or equivalent docs you’ve put together. 

P.s. I'm running a local Ollama instance on an older RTX destop with 64GB of RAM and it's still paaainfully slow once youre used to having a cloud connected datacentre running Opus/Fabel!
Hope it's a big birthday cause you'll need a LOT of compute :-)

My experience is similar to yours. I started before you, and back then there was actually no client app support for connecting to the MCP server because MCP was only a couple of weeks old—it mainly existed in documentation and had initial library implementations for both the server and client. Because of this, I looked for an AI client supporting MCP and found Cherry Studio. I will write some posts about the parts you are interested in, but sometimes it is better to struggle and learn from your mistakes.

Thank you.  I've been running it live today and very please with how it is working.  I still need to understand much more but bfexplorer is vey flexible.  I've not looked at running local LLM yet.  I have Ollama downloaded but my focus was in having a bot running first.  I may need a PC upgrade with LLM.  I will see.  Fortunately it is my birthday soon :-).  I tried a nimber of ai client applications.  Most I found disappointing with repetitive errors.  Claude was easiest for me to work with but I had to be very specific to ensure it did not "forget" what I was trying to do.  I made claude read and summarise information for it to use as quick reference. I learned to be very specific with wording on messages for what I wanted and also when starting a new task I asked clause to commit to memory any background research or help documents before starting.  If I started again it would take me a lot less time to have the same results. I will advise on local LLM when I have investigated.

I just posted "Wiring an AI agent into a Betfair trading desktop" so if still not working let me knoe and I will see if I have anything to help.

I am highly impressed with how you integrated bfexplorer agentic support using the Claude LLM and your strategy execution pipeline that incorporates market data context processing, as any robust strategy must be driven by verifiable data.

Have you also experimented with local LLMs or different AI client applications/harnesses?

A simple test to verify that the bfexplorer MCP is running correctly involves using the 'modelcontextprotocol/inspector.' To do this, run the following command in your terminal:

npx @modelcontextprotocol/inspector

After setting up the bfexplorer connection, you will be able to test all available tools without requiring an LLM.

Because you used Claude Desktop, I noticed it has a poor user experience for setting up the local MCP server. Please refer to this document, which shows how to set up the MCP server on client apps.

Bfexplorer MCP Client Application Setup

Thank you for letting me know about the issues with my website.

I run my site on a Windows Virtual Private Server hosted on FORPSI, which costs me approximately 500 Euro annually. This is currently a side project for me, and since I only have around eight paying users, it does not make economic sense for me to invest in something more reliable. Everything has an associated cost; therefore, I currently do not have a security certificate for my website or a separate certificate for the desktop app.

I plan to move the app to the Microsoft Store, as this would potentially provide a free certificate, but I am unsure if a betting app of this nature would pass their Terms and Conditions (T&C). For now, I have removed the installation file for bfexplorer from my server because its presence could cause server timeout errors when more people attempt to download it.

Of course, no service can be 100% reliable; occasionally, the cloud provider restarts servers or network issues can affect web app accessibility. The server also requires periodic restarts, especially after critical updates that necessitate a reboot.

I personally use the Chrome web browser, and I have not encountered any issues with my website while using it. I also run several related sites for bfexplorer, including my main site at Bfexplorer - Betfair Trading Software, which I created for my Reddit community, Betfair AI Trading: https://www.reddit.com/r/BetfairAiTrading/. Furthermore, you can view my supporting GitHub projects here: my community repository is https://github.com/StefanBelo/BetfairAiTrading and the Bfexplorer BOT SDK project is https://github.com/StefanBelo/Bfexplorer-BOT-SDK.

Install from Bfexplorer Preview version. 

This is going to take time for to understand and work through Stefan, and in fairness as my content is not yet useful I don't deserve free use so I have subscribed to the basic model.  Three markets will help me as I need to split feeds into a minimum of three distinct market types.  Anyway, I will document my journey as you suggest and then we will see. My expertise is commercial contracting so I have a steep learning curve.  

I have been coding and trading on Betfair since 2007; I am currently 60 years old. When I first built my website, I introduced a Community Program offering free professional subscriptions to our most valuable community members. Do you know how many users took advantage of this general offer? Zero. Why? Because that is the nature of bettors on Betfair.

If you want to participate, you must produce readable content detailing your journey and genuinely welcome it here. In return, you will receive free Betfair API access. Since you already have experience with AI models, I suggest writing about that process—specifically how you use the bfexplorer agentic app with AI.

Furthermore, if you have a computer equipped with some VRAM/GPU, you can run AI models locally. For example, I successfully tested running the Gemma 4 LLM even for prompts similar to those other users have asked.

Time to process after market change

You need to know how to use local LLMs and, crucially, write effective prompts.

Thank you for your response.  It will help me in the direction I go.  I am not an expert on this or a programmer (I'm a retired 68 year old having fun) but I have been trying to find help through AI including including a claude subscription as this was advised as best to use.  There are betfair apps that I could make some direct links and potentially two that I have found that could potentially use the marketid information however even the ones that suggest that they can do this have some limitations.  The initial AI advice was to use an excel dashboard which eventually was finished after many corrections however the tasks I need done are linear and because of this excel was slow.  My major focus now is learn how to use bfexplorer (so looking at a lot of youtube videos) and I will look at the script also. I have a strong impression that bfexplorer is very configurable.  Once I have that working I am happy to move to a pro subscription.  Trying to build my own betfair app is much more about having something to do and to keep my brain active.  It's nothing to do with money.  If I have something I am happy with I will pay the the £499 for for individual access and to support your continued work.  The reason I am here is to become more informed about these topics.  I know I looked a long time ago at bfexplorer but did not have the time to understand everything. All the best.

I do not understand your use case; as for why you need to open Betfair markets—if it is for betting purposes, those actions can be fully automated in bfexplorer without any programming knowledge.

Yes, there are many ways to achieve this using simple code, and I have provided several such examples on my GitHub.

The simplest option is a bfexplorer Console script; open it in the bfexplorer Console and click the "Execute" button on the toolbar.

BfexplorerConsoleExample.fsx

Since bfexplorer is an agentic app, you could simply set up an MCP connection with your AI harness and prompt bfexplorer: "Open the following market IDs: 1.261142591, 1.261142708."

Well, I understand your motivation for getting free access to the Betfair API through my generous community program, but you should be at least somewhat informed about these topics.

Open Claude Desktop. Navigate to File > Settings, then open Connectors and click Add. Select "Add custom connector," name it "bfexplorer," and for the remote MCP server URL, enter: http://localhost:10043

To be honest, it seems like you are missing something: in the era of AI, you can simply ask any AI tool how to add a feature. Many AI-driven applications do so automatically; if they cannot find the necessary information in your environment, they will prompt you and configure everything themselves.

What does that mean for you?

I don't think you’ll be able to utilize bfexplorer’s agentic support effectively, so you should just stick with using the app the traditional way.

Have you got a guide on how to setup and use either Copilot or Claude?

When you activate "Use web page login" in Application/Settings, it means that when you log in to Betfair via the bfexplorer app, a dialog will appear displaying the official Betfair web page login. This allows you to complete the second step of your two-factor authentication (2FA). To be frank, I have not tested 2FA yet, but the standard web page login works. My only hope is that after a successful login, the 2FA process redirects back to the same webpage, as bfexplorer needs this location to read the session token and authorize any of your Betfair API operations.

Set up bfexplorer agentic support in your AI chat tool, and then simply ask bfexplorer your questions.

Can you advise how to set up this strategy? Say for example I want to monitor in-play horse racing and place a bet on a horse only when it moves to price X, assuming the starting price was Y or higher?