• What it was — a multi-model LLM workspace: one conversation that switched between OpenAI, Anthropic, Gemini and xAI mid-thread without losing its context.
• My role — everything.
• Status — shelved deliberately, and kept on this site as the decision rather than the product.
• What survived — the hierarchical context compression, which now runs Eve's rolling summaries in DrawEvolve.

What it was

A multi-provider LLM orchestration platform: routing across model providers, with hierarchical context compression so long sessions stayed usable without blowing up token costs.

What it did

Live notes, written as the conversation happened. A rail beside the chat took notes on the session while it ran, in snapshots, pulling out key topics and keeping a running summary of what had actually been discussed. A read on the conversation, produced while you were still in it.

Commands it wrote for you. Lynk generated invocable commands out of the conversation itself. You could call them in-chat to perform a function instead of describing it again in prose every time.

Four models from four providers, inside one conversation. Claude, OpenAI, Gemini and Grok, switchable mid-thread, with context carried across the switch. One conversation that changed engines under you and remembered everything that had been said to the previous one.

Summaries used as memory. As a session grew, Lynk summarized its own older turns and then referenced those summaries instead of replaying the full history. Hierarchical compression, to spend fewer tokens and hold more session. The compression is the part I'd still defend. Long-running LLM sessions degrade in predictable ways, and summarizing older context while keeping recent context intact was a genuinely reasonable answer to it. The same idea, scaled down, is what runs Eve's rolling summaries in DrawEvolve today.

All four of those shipped before the major providers shipped anything comparable. The reading was right. Holding the position turned out to be a separate problem, and that is why the project is here.

Why I killed it

The unit economics didn't work for one person. Every user session cost real money against provider APIs, the value proposition was thin against first-party tools that had the same models and more engineers, and I would have been maintaining provider integrations indefinitely while doing it alone.

I shelved it, kept the repo, and moved the parts worth keeping into DrawEvolve.

Why it's on this site

Because knowing when to stop is a skill, and I'd rather show you a project I ended deliberately than pretend everything I've started is still alive. The compression architecture found a better home. I planned a different ending and I'll take this one.