From Assistant to Action: Highlights from the New io.Intelligence Release

This is not just a product update. It represents a complete, production-ready stack for firms that are ready to move from experimenting with AI to actually deploying it in the hands of users.

That was the focus of a recent interop.io webinar, “io.Intelligence New Release,” featuring Bob Myers, Chief Product Officer at interop.io, alongside Kalin Kostov, Engineering Lead for io.Intelligence, and Krasimir Krastev, Director of AI Solutions.

AI that understands where you are – not just what you ask

One of the most striking moments in the session was when Krasi opened the demo by simply greeting the assistant. Within seconds, it identified itself as a trading assistant embedded in the workspace – aware of the applications on screen, the context of an OMS, and the specific security being worked on. No configuration required mid-session. No retrieval queries. It just knew.

This is the promise of Working Context, one of the foundational layers of io.Intelligence. Rather than requiring users to explain their situation on every prompt, the system tracks information already being shared by applications in the io.Connect platform – channel context, workspace context, application context – and makes that available to the language model automatically.

Crucially, nothing is exposed to the LLM unless explicitly configured. Firms define a schema that specifies exactly what data is tracked and where it comes from. As Kostov put it during the architectural walkthrough:

“You can be one hundred percent certain that you are in control of what is made available to the LLM.”

The result is an assistant that can interpret relative language naturally. Users do not need to reference client IDs or instrument codes. If the workspace already knows who the client is and what security is in focus, the assistant knows too.

A full stack, not a point solution

The Q1 2026 release builds significantly on what shipped at the end of 2025. The earlier release established the Working Context module and the MCP Core tooling layer – giving firms a way to connect io.Connect platform capabilities and application methods to any AI assistant they were already running.

The new release adds the remaining layers to make that a complete, self-contained system:

io.Assist is a ready-to-use chat interface that firms can deploy immediately or use as a starting point for their own design. It is intentionally lightweight on the UI side, because the intelligence sits elsewhere.

AI Web is the front-end runtime module that handles the actual AI logic: connecting to agentic back ends, hosting multiple MCP servers, and carrying the system prompts that make Working Context and system tools work together efficiently. It is framework-agnostic, which means firms that want a completely custom interface can still use it.

io.Intelligence Server provides the back-end orchestration layer, built on Mastra AI, a TypeScript-first framework for agentic systems. Critically, it communicates over AG-UI, a protocol that describes how front ends and back ends of agentic systems should talk to each other. That means io.Intelligence’s front end is not locked to a specific back end – any AG-UI-compatible system can plug in.

Bob’s summary of the combined effect:

“You’ve turned essentially one assistant and LLM – or multiple LLMs – into a universal remote control for your enterprise desktop.”

Governed, observable, and explicitly permissioned

A consistent theme throughout the webinar was the contrast between io.Intelligence’s approach and the broader category of “computer use” AI – systems that observe everything on screen and attempt to take action based on what they see.

Bob was direct about why that distinction matters:

“Everything we showed you today are explicitly permissioned capabilities that an AI can take on your user’s behalf, and you control that.”

Every tool the assistant can call is a method or intent registered from an application running in io.Connect. That means firms decide in advance what the assistant can and cannot do. An assistant can open and populate a trade ticket. Whether it can actually submit that order is an implementation decision, not something the AI determines for itself.

And because everything runs through the io.Connect platform, observability is built in. Firms can see exactly what happened, which tools were called, and what data flowed through the system – the same kind of transparency the platform already provides for interoperability workflows.

MCP Apps: bringing interactive UI into the conversation

One of the newer additions demonstrated in the session was support for MCP Apps – a standard that allows AI assistants to render interactive UI widgets directly in the chat canvas, rather than responding with text alone.

When io.Assist is running inside a workspace, rather than rendering widgets into the chat canvas – where they compete for space with the conversation – the system can push them into the workspace itself, using available screen real estate more effectively. Widgets can also communicate back to the assistant, triggering further actions. In the demo, clicking a “book” button inside a widget sent an instruction that dismissed it from the workspace automatically.

Meeting firms where they are

Not every firm is at the same point in their AI journey, and the session addressed that directly. For firms with no AI assistant yet, the full stack provides everything needed to start from scratch. For firms already running an assistant, io.Intelligence can extend its capabilities with richer tooling and workspace awareness. For firms with mature assistant deployments ready to introduce autonomous agents, the framework supports that too.

The underlying principle is the same in each case: the investments firms have already made in io.Connect – application directories, interop methods, workspace management – are the foundation AI needs to be useful. io.Intelligence does not require a rebuild. It adds the orchestration layer for AI to take action.

Myers returned to this point near the end of the session:

“If you’ve made all these investments in interoperability, you may be further ahead than you realize in terms of introducing useful AI capabilities to your users.”

What comes next

Looking ahead to Q2 and Q3 2026, the team outlined three areas of focus.

The first is token efficiency. Running large language models at scale across a business day has real costs, and io.Intelligence is addressing this with a “code mode” capability – allowing the assistant to write and execute scripted programs rather than invoking tools repeatedly. Early research suggests this could reduce token usage by as much as 80 percent for certain workflows.

The second is context window management. Supporting users across a full working day requires careful stewardship of what the LLM holds in context at any given moment — compressing, summarizing, and maintaining relevance as conversations evolve.

The third is observability tooling, extending the io.Insights product line into the AI layer so that firms can understand not just what their applications did, but what their AI systems did alongside them.

Further out, the team is focused on ambient agents: systems that respond to notifications and events in the background, operating alongside human users and conventional assistants rather than replacing either.

The takeaway

The Q1 2026 io.Intelligence release closes the gap between a compelling demo and something firms can actually ship. The working context layer, the MCP tooling stack, the front-end runtime, the agentic back end, and MCP App support are all generally available.

For firms that have been watching AI developments carefully and waiting for the right moment, the infrastructure is ready. The question now is where to start – and for io.Connect clients, the answer is closer than most might expect.

If your team is ready to explore what this looks like in practice, the io.Intelligence documentation is available now.