GitHub Copilot Dev Days
Welcome - Precious Ajuru
Precious welcomed everyone to their GitHub Dev Days event in Newcastle upon Tyne on 31st October as part of the GitHub and Microsoft events happening around the world. They also run the North East Azure Hub events which have talks and many opportunities to connect about Azure along with related topics. Microsoft Ignite is happening in San Francisco in November which they will be attending in person, but it is also possible to attend virtually for free, to find out more and attend online during the event can visit ignite.microsoft.com. Also there was a chance to build a Space Quiz with a GitHub Copilot app workshop which can be found at github-samples.github.io/copilot-workshops/first-steps/copilot-app.
Take Stock of Model Context Protocol with .NET - Peter Bull
Large language models can be powerful, they can generate text or code, but they can't do anything in the real world, without a bridge between models and reality. Tools perform actions on your behalf to invoke functions with named operations along with defined inputs. Resources deliver structured data to context of large language models with Markdown or other formats, from a named data source with description and schema defining the structure. Prompts control output from large language models to provide templates with reusable instructions that can shape a large language model's approach to tasks.
Model Context Protocol, or MCP, is the standard to bridge AI apps from large language models to reality with external systems, it was created in 2024 by Anthropic with many MCP servers being implemented and used daily by developers such as for Jira, Azure DevOps and more. Servers implement Model Context Protocol to expose actionable tools, define data resources and provide reusable prompt templates to clients. These clients such as GitHub Copilot can leverage Model Context Protocol servers to take advantage of their functionality. To find out more about Model Context Protocol or MCP then visit modelcontextprotocol.io.
Model Context Protocol is an evolving standard, with support for Skills coming soon, these provide reusable capabilities to enable AI clients to perform specialised tasks currently supported today in many clients with local SKILLS.md Markdown files, which define when they should be used and contain instructions or workflows that can combine actions from tools or knowledge from resources and define expected inputs or outputs. Servers will be able to add to the capability of local SKILLS.md files with Skills over Model Context Protocol with updates to libraries to enable servers to support this along with clients will be coming soon.
Model Context Protocol extends the ability of large language models with actions that can use natural language requests such as finding products or updating stock. You can build for Model Context Protocol in a variety of languages and frameworks, such as .NET which is Microsoft's open-source cross-platform runtime that supports desktop, mobile and web for modern applications and C# the programming language used in .NET where there is an official C# MCP SDK to expose tools, resources and data. You can connect a client supporting Model Context Protocol such as GitHub Copilot which can use the appropriate tools and resources to present results. You can leverage Model Context Protocol in Visual Studio 2026 on Windows which supports the broadest range of features for tools, resources and reusable prompts or take advantage of most features in Visual Studio Code and GitHub Copilot app. You can find out more about .NET at dot.net, GitHub Copilot at gh.io/copilot or Visual Studio 2026 at visualstudio.com.
Model Context Protocol can be used through a variety of developer tools such as Visual Studio 2026 which can use GitHub Copilot as the client to connect natural language requests to a server implemented in .NET and the C# MCP SDK, available as a seperate template in .NET 10 but built in with the upcoming .NET 11, using MCP over HTTP. Server can provide a variety of tools to enable a client to find or create products or update stock, sales and get sales history that uses a database for storing operations along with code providing deterministic validation of actions so invalid actions such as creating negative stock or cost situations cannot be performed. Resources can provide clients with a sample catalogue, stock examples and expected results stored in embedded Markdown, which can also be used to for prompts which provide reusable instructions to guide how client can approach or perform tasks. You can check out a demo using Model Context Protocol with .NET using the C# MCP SDK at github.com/rogueplanetoid/takestockofmcpwithdotnet.
Agentic Commerce - When the buyer is an AI - Godwin Jimmy
Godwin Jimmy has been a CTO for a Fintech company for eight years and mentioned that finance keeps evolving and means we makes payments and seeing definition type change from end users to agents and they started to explore opportunity which is to vibe finance, are exploring as much as they can to see what catches on at the earliest stages and have one of the earliest agentic wallets on the Apple App store but had to define what that means but managed to get on the store recently. What happens when user is an agentic agent?
Most of us have seen the flying car, this is the kind of thing we will see going forward that things in movies are practical in real life, imagine someone sends a message about getting pizza so how do you handle this where can use Claude to do this but it would reply and ask if you want it and then once done can approve the payment to, in less than twenty minutes something happened and a transaction was made and for most of us would not have to know and a decision was made and transactions sorted, this where we are going but there are a lot of independent moving parts. Agentic finance is not just getting money from bank account but be able to give AI a tool to be able to spend money.
An agent finds a tool, pays for it and uses it for you - Agentic Commerce software agents buy on a person's behalf inside limits that person sets, you set the task and spend limit, agent picks the tools and the tool asks to be paid and then answers. The limit is important the scenario depends on what you are getting, pizza over a car so this important to distinguish, the task you define is defined and the wallet has its own spending limit. Tools from Model Context Protocol are like this, they spoke about the agentic era is here and early in MCP was it a fad and now there is a handful of tools available such as emails or even a virtual mobile phone or where could make a actual phone call from an AI, the agent picks a tool it needs to make something happen such as buying from Amazon and make a vibe purchase. There are 600,000 tools but only 4% are paid tools but is quite early so most people are doing things for free, but the actual value is data itself so the price on the data is valuable so when want to get a projection of data you need historical data or for example you could ask for where to find the best price for something.
There are 620 million projected transactions and 40% of online shoppers use AI to research the most recent purchases which is why ChatGPT brought in Ads to show suggested products, $19.7 billion of AI platform transaction sales and AI agents are paying for MCPs, APIS, research papers, products and tools so how are you positioned for this? Should see a 4 to 6 % adoption rate, where Shopify has provided an MCP where can get AI agents to do more things for purchases, early-stage adoption will be selling of data or micropayments, no one is going to purchase anything large but if a small 50p or £1-2 payment when would be more willing to do that.
Your MCP / AI tool can be a blank slate as an AI agent can by a buyer, if you build tools like an ERP based MCP could build tools, sell per call, that charge per call without building a checkout, key storage or billing where each tool gets its own address which works in Copilot, Cursor, Claude or any MCP client. If you use agents, you let it pay, on a leash where the agent can buy a tool mid task instead of stopping, a per payment max and daily cap you can set and can approve or decline each payment or revoke the key any time. We should understand agentic finance and commerce doesn't mean giving everything over can have guardrails so an agentic wallet is something that the AI agent can access not your own wallet, bank account so in case the AI for any reason anything happens any risk is restricted to that wallet itself. With their system you generate a vibe access payment token, where can have this issue to an AI agent and know that the client has made that payment where each one of these has a per transaction or daily limit and it cannot spend more than that. You can also add trusted sellers such as Amazon or Domino's Pizza and block unapproved ones where took the OTP approach to validating transactions but using face or fingerprint so that cannot be performed by an agent.
Today's flow stops the moment money is involved, find the tool, create an account, add a card at checkout, create an API key, paste the key into config, call the tool where steps between first and last need a person so the agent stalls. Majority of users are humans but the rest will be AI agents but most systems when AI agents try to interact if they managed to go through things and get to a subscription page you optimised it for humans not an AI agent so it gets there it stalls and needs a card to make a payment and AI agents when provided with card details will reject this so automatically you have lost that customer. One of the things they did where they took am old request type of 402 / Payment Required which turned into X-402 and this entire protocol is open source so took this and added into their system where have an SDK and you put this into your application. For most checkouts have internally have the capture card but the agent gets a 402 page and knows it needs to make a payment, then sees if there is a wallet and then sees the amount and then if within limits it can make the payment, otherwise the flow breaks. Are on the Stripe programme where they can charge an agentic charge for micropayments a single Stripe payment is 0.40c but can round up numbers into thousands requests to make it easier but maybe you only want one so can support Crypto payments.
Three assumptions an agent cannot meet, 1. Is that checkout is a form - email card fields and captcha built for a person in a browser and an agent in Copilot has none of these. 2. The API key is the wallet - a pasted key usually opens the whole account and there is no per call limit the agent must stay under. 3. No price the agent can read - nothing in the response says what the next call or how to pay for it. If fix all three the agent can pay without a human in the browser, while the human keeps the limits. You need to make sure have cases covered and guardrails, if you have a payment gateway you are listening to success or failed if you listen to a webhook from Stripe that's not your problem that is theirs. Have seen agentic wallets on a CLI basis where you request a key which can be good but it also not good as this key is accessible to anyone and everyone with that access, when see agentic wallets and the key is the wallet this can be risky.
Ask, pay, unlock, so call is where agent calls tool like any MCP tool, asks to be paid with free calls used up and then 402 payment required plus a vybe.pay sign in and pay link, approve where agentic wallet MCPs pay inside your spend limit or ask to approve and then the same call again the retry carries the token and the tool answers. First step is a URL in your MCP config where needed and then the tool asks to be paid with HTTP 402 where get this response and then can tell the client which sign in to run and a pay URL for a link to a person can open. Did see cases where the wallet is only charged if there is a 200 response so there is no payment taken for failures. They do have a counter for how many requests are made so can purchase a batch of requests.
Vybe fits in where on by side use the Vyve Pay MCP with an agent walled with spend limits and approve / decline, Buildr for building a tool in plain English where this is hosted with own MCP address, then get paid along with web games, but you can create a paid tool where it is free to start and then tools ask to be paid and then agents can unlock this. You can find out more about Vybe at vybe.finance.