AWFlow wants to be the OpenRouter for AI that runs in your browser

The goal. Say it slowly.A 3-second opener before the first chapter.The goal. Say it slowly.A 3-second opener before the first chapter.

Locked in. Pick an engine, it controls your models.A model prepared for one engine will not load in another.Three plug shapes. All carry electricity.Different form factors, same function — the analogy maps to browser AI engines.Same thing inside. Same AI math, three different formats.WebLLM uses MLC-compiled weights; TF.js uses model.json; transformers.js uses ONNX.Adapters needed. You pick a format and stay with it.No shared on-device model format exists in the browser today.Locked in. Pick an engine, it controls your models.A model prepared for one engine will not load in another.

A router hides this behind one interface.Caller describes a task; the router picks the engine and model.Three tasks your app might need.Chat, vision and speech currently require different engines.Three engines to initialise and maintain.Each engine has its own load, initialise and infer API.The work fans out to the right engine.Each task goes to the engine whose models can do it.A router hides this behind one interface.Caller describes a task; the router picks the engine and model.

Swap the model without rewriting anything.If the model is a dependency node, changing it is a one-step edit.Know the task. Chat, vision and transcription each need a different engine.A task registry maps each job type to supported engines and model families.Know the hardware. A model that will not fit is useless.Estimate memory before offering a download; show a device-fit warning when uncertain.Download the right format. Each engine needs its own.MLC builds for WebLLM, model.json for TF.js, ONNX for transformers.js.Route by job, not by engine name.The caller describes the task; the router resolves which installed model handles it.Swap the model without rewriting anything.If the model is a dependency node, changing it is a one-step edit.

123

1. 1Grouped by taskChat, embeddings, vision, speech-to-text.
2. 2One card per modelEngine, size and device fit.
3. 3Shared storageOne download, used everywhere.

This is the router made visible.The page is the interface between you and the three engine layers beneath.Tasks first. You see jobs, not engine names.Models for WebLLM, TF.js and transformers.js are shown in one list, grouped by task.One card per model shows what you need to decide.Size, estimated device fit and install state. Engine is an implementation detail.One install, everywhere. The assistant, agents and workflow nodes share it.A model installed on Local AI is usable by any Web LLM node without re-downloading.This is the router made visible.The page is the interface between you and the three engine layers beneath.

Swap the model and the job still works.The resolution layer finds the new model; the workflow node stays the same.You send a job. AWFlow sends it to the right engine.The task name resolves to an engine and a model at runtime.The engine is picked automatically. You never name it.Task-to-engine resolution is done by the Local AI layer.The result comes back. No engine name in it.The output is the same regardless of which runtime handled it.Swap the model and the job still works.The resolution layer finds the new model; the workflow node stays the same.

123

1. 1On this deviceFree, private, offline once installed.
2. 2Or a cloud providerYour own key if a job needs more power.
3. 3Start chattingSame interface either way.

The choice is yours from the first screen.Users are not locked to a model or an engine at sign-up.On this device means a browser AI engine does the work.Selects an installed Web LLM model from the Local AI page.Or a cloud provider with your API key.Chat with OpenAI, Anthropic or Google, called directly from the browser.Same interface either way. Change the model, the experience stays.The assistant does not expose which engine handles the request.The choice is yours from the first screen.Users are not locked to a model or an engine at sign-up.

The fact. Each website or extension keeps its own copy.Cache Storage and IndexedDB are partitioned by origin; no cross-origin model registry.The myth. One download, one gigabyte.A system library is installed once and shared by every process.The fact. Each website or extension keeps its own copy.Cache Storage and IndexedDB are partitioned by origin; no cross-origin model registry.

In-browser AI is at the same stage. The path is clear.Three islands, no common format — the same shape cloud AI had in 2022.The cloud had the same mess. Routers cleaned it up.Before routers: per-provider SDKs, per-provider billing, per-provider migrations.In-browser AI is at the same stage. The path is clear.Three islands, no common format — the same shape cloud AI had in 2022.

Summaries lead. Chat close behind. (Example results.)Example only.Your turn. What is the job you need done?Example results, labelled on screen; no data is collected.Summaries lead. Chat close behind. (Example results.)Example only.

Three. The hard problems are above any single product.Per-origin storage and engine-specific formats must be solved at the web platform level.One. WebLLM, TF.js and transformers.js are islands today.Each engine expects its own model format; no cross-engine format exists.Two. AWFlow sits in front of all three.Install on Local AI; the assistant, agents and workflows all use the same model.Three. The hard problems are above any single product.Per-origin storage and engine-specific formats must be solved at the web platform level.