The gap between a clever prototype and a companion you'd ship
Most AI prototypes die in the space between the demo and the device. A companion that looks brilliant in a notebook has to survive contact with real hardware: limited memory, no guaranteed network, and a user who expects it to behave the same way every single day. That last mile is where teams burn weeks wiring together ad-hoc scripts, spreadsheets of test prompts and manual release checklists.
Edge Studio exists to close that gap. It is the workbench behind LyboAI Edge, our on-device AI companions, and it takes one opinionated position: the path from idea to device should be a single pipeline, not a pile of disconnected tools. Everything in this post describes how that pipeline works today — Canvas for designing behaviour, projects for keeping the work organised, evals for proving it, and signing for shipping it.
Canvas: design the behaviour, not just the prompt
Canvas is where a companion starts. Rather than a single wall-of-text prompt, you lay out the companion's behaviour as connected pieces: its role, the tools it can call, the guardrails it must respect, and the handoffs between steps. Because the pieces are visible, the whole team can reason about them — a product owner can read a Canvas the way they'd read a flowchart, without needing to parse prompt engineering folklore.
Each Canvas lives inside a project, which is the unit Edge Studio organises everything around. A project holds the Canvas, its versions, the eval suites that test it and the builds produced from it. When you change the companion's behaviour, the project keeps the history, so 'what changed between last week's build and this one' is a question with an actual answer.
Evals: prove it before anyone installs it
The honest truth about AI companions is that you cannot eyeball your way to reliability. A change that fixes one conversation quietly breaks three others. Edge Studio treats evals as a first-class part of the pipeline rather than an afterthought: you define suites of scenarios the companion must handle, run them against a candidate version, and get a pass/fail picture before a build ever leaves the studio.
This matters twice as much on-device. Once a companion is running on someone's hardware, you don't get to hotfix it mid-conversation the way you might patch a server. The eval gate is how we make 'ship it' a decision backed by evidence instead of vibes. Our VALUE-AI framework shapes what those suites cover — usefulness, reliability and honest behaviour under pressure — so teams aren't inventing their quality bar from scratch.
Signing: the moment a build becomes trustworthy
When a version clears its evals, Edge Studio signs the build. Signing is the studio's way of saying: this exact companion, with this exact behaviour, is the one that ships. The device-side runtime checks that signature before running anything, so what your users get is what you tested — not a near-relative of it. It also gives you a clean audit trail: every signed build traces back to a project version and the eval results that let it through.
This is deliberately boring engineering, and that's the point. On-device AI asks people to trust software running close to their lives, often with personal context that never leaves the hardware. That trust should rest on a verifiable chain, not on a vendor's word.
Where it sits in the wider platform
Edge Studio is one piece of a deliberately small stack. Lybo OS provides the environment companions run in on the device, the Agents Platform handles orchestration when companions need to work together or reach out to other systems, and Edge Studio is the front door where behaviour gets designed and approved. The seams between them are boring on purpose — a signed build from the studio is exactly what the runtime executes.
We should be straight about maturity: Canvas, projects, evals and signing are the working core of Edge Studio today, and we're continuing to deepen the eval tooling and orchestration story rather than bolt on new surface area. If a capability isn't in this post, treat it as roadmap, not product.
Try the pipeline on your own idea
The best way to understand Edge Studio is to push a real idea through it: sketch a companion on Canvas, write five eval scenarios you'd genuinely worry about, and see whether your design survives them. Most people discover their second version is meaningfully better than their first — which is exactly what a pipeline is for.
Explore Edge Studio, LyboAI Edge and the rest of the platform at lyboai.app. If you're weighing up on-device AI for your product or team, start there — and hold us to the same standard the eval gate holds our companions to.