What if an AI could extend its own capabilities automatically? Imagine if you asked an AI to generate a new fully integrated company e-commerce platform. Most AIs today would provide you with a list of tasks that are needed but that would be it. However if instead the AI asks you for permission to download connectors for and starts connecting to Salesforce, SAP, BigCommerce, Google AdWords, Stripe, … in order to create your fully integrated e-commerce platform featuring online advertisement [AdWords] tracking sales leads [Salesforce], selling products online [Bigcommerce], handling operations and accounting [SAP], taking payments [Stripe],…
For this reality to be possible, we need to be able to securely extend AI’s capabilities. So how can we do this?
Press enter or click to view image in full size
What is MCP and why should I care?
Antrophic last year open sourced MCP. OpenAI, Microsoft,… last month all launched products supporting the MCP standard.
MCP is the USB connector of AI: enabling any tool, data source and workflow to be plugged in and extend AI into anything. However MCP is far from Enteprise-Ready, that is until now… [keep on reading :-) ]
Thousands of MCP servers are popping up everywhere. LLMs can now be integrated into anything.
The MCP reality today
The hype machine will make you believe that MCP is going to solve all the business problems that client/server, big data, cloud, api gateways, Kubernetes, micro services,… were promising to solve.
However if you analyse what is really available today then you find a lot of JavaScript and Python scripts that are started up as a server, take some environment secret and which connects to often critical systems. These “MCP Servers” lack any enterprise-readiness, no observability, no security, no scalability, no nothing.
Rethinking MCP for the enterprise
MCP connectors need to be reliable, secure, scalable, observable, compliant, easy to deploy/upgrade,…
Downloading a random script from the Internet on top of a Python or JavaScript version manager which can spit out random problems is the opposite of Enteprise-ready.
We need connectors to be packaged in a way that we know will always work. Connectors shouldn’t bring security and compliance issues. They should handle secrets securely and most important of all we want to have a good night sleep. Connectors should be the most boring aspect of our AI and IT strategy. They should just work and never gives us problems.
Steal another standard with proud
The IT industry has many solutions if you know where to look. One of the most amazing new technologies is WebAssembly or WASM. WASM allows you to package hostile and badly written code and run it securely. If you used Figma, Google Earth, … then you used WASM. WASM is a standard which browsers use to run programmes written in very efficient languages like C, C++, Rust, Golang,… securely to optimise complex websites.
The WASM standard has been embedded in all major browsers for years. More recently WebAssembly System Interface or WASI has come along which allows WASM to escape the browser and be used as a micro container with a standard interface everywhere. WASI enables a WASM to securely interface with very specific outside worlds services, e.g. a WASM can connect to a certain URL but to nothing else. This secure containerisation is what makes it an ideal technology to package MCP connectors.
WASIX-MCP
By defining a WASI MCP interface which exposes the MCP tools, resources, prompts,… but also allows external http connections as well as managing secrets [e.g. API keys, oAuth refresh tokens,…], we are able to lay the foundation for enterprise-ready MCP connectors.
So by downloading one WASM container and setting up the secrets it asks for, we are able to dynamically extend the capabilities of any AI, e.g. LLM, agents, Agentic,…
WASM can be signed, so security, compliance, scalability, observability,… can all be certified before deployment. Upgradability is as easy as dropping in another WASM.
But here is the real kick. The first WASIX-MCP connector: the weather API connector, comes in at around 200KB in size. Most photos and images you download on the Internet are 5 to 10 times larger than this connector.
So security, provenance, compliance, observability, scalability, stability, upgradability, performance,… all in a couple of hundreds of kilobytes. Drop this into your AI solution and it allows your AI agents to get the weather. Update the code to connect to anything else and you can integrate your AI Agentic army with the world.
Conclusion
The WASIX-MCP is one small (200KB) file for AI Agents but a giant leap for Agentic AI. Any AI will be able to download in a few megabytes hundreds of connectors and extend itself dynamically in a secure way.
If you are working on AI Agents and Agentic solutions for your business, please be sure to reach out.