Think inside your AI world.
Does Unl use up my context window, or distract the model?
It uses a small slice, on purpose: only the decisions that bear on what you just asked, each with its reason, and nothing your AI already has. Here is what that costs, measured.
Unl does not paste everything you have decided into every prompt. Before each turn, Agent Unl chooses only the decisions that bear on it, each with its reason, and leaves the rest out. It shapes what it sends for the model reading it, and it does not resend what your AI already has in the conversation. In Claude Code a typical serve is about 2,000 tokens. On real fleet runs, Unl cost 15% more tokens, and the agents acted within what had been decided on 70 of 90 tasks instead of 61.
What Unl sends, and what it leaves out
Unl holds everything you have settled, and sends very little of it. Before each turn, Agent Unl reads what you asked and picks the few decisions that bear on it. The rest stay held, one call away if the work turns that way. A decision that does not bear is not sent, so it is not there to pull the model off course.
A decision is not sent again while your AI still has it. When your AI already has it in the conversation, later turns point back to it instead of repeating it, so the cost does not pile up as a long session runs on.
How much of the window it takes
In Claude Code, the typical serve is about 2,000 tokens, and nine in ten stay under about 2,100. That is measured across 137 serves from 25 to 27 September 2026, from Unl's own serve log. Against a 200,000-token context window, it is about one percent.
In chat apps such as ChatGPT and Claude on the web, serves run larger today, because the model there reads Unl through a connector. Unl has started shaping those serves to each model, and that work is under way.
The cost, measured on real fleets
We ran the same 45 tasks across 9 projects through two agent fleets, with Unl and without it, under a protocol hashed before the first run. With Unl, the fleets acted within everything decided on 70 of 90 tasks instead of 61. The cost was 15% more tokens, Agent Unl's own work included, under a limit of 25% set in advance. The baseline was already careful: each agent was told to read a well-kept decision log.
A second run, with Unl as it ships today, came in at 67 of 90, with fewer decisions crossed, no added time and 19% more tokens. Both write-ups are public, with the protocol and the numbers behind them: the result of record and the re-run.
Does it distract the model from what I asked?
Your instruction leads. Unl's decisions arrive as labelled background beside your message, never in place of it, and each carries its reason, so the model can see why it bears and how much weight to give it.
The fleet test measured focus directly. With Unl, the finished work served the project's stated aim better, by 0.56 on a five-point scale, not worse.
Unl spends a small, measured slice of the context window on only the decisions that bear on the turn, and in return agents act within what was decided more often.
Reads through Unl arrive with measured context — in the presence of the decisions you’ve already settled. The reach lane is live: one box, paste anything. If it speaks MCP, Unl can reach it. Readings arrive unprompted, the data beside the criterion; Unl is a courier, not a warehouse, and keeps only your keys and the frame.
Read further
Questions people ask
How many tokens does Unl add to a turn?
In Claude Code, about 2,000 for a typical serve, and nine in ten serves stay under about 2,100 (137 serves, 25 to 27 September 2026). Chat apps receive larger serves today, and Unl is shaping those to each model.
Does Unl resend the same decisions every turn?
No. Once your AI has a decision in the conversation, Unl points back to it instead of sending it again, and serves only what is new or newly bears on the turn.
What does the extra 15% of tokens buy?
In the fleet test, agents acted within everything the person had decided on 70 of 90 tasks instead of 61, and the finished work served the stated aim better. The re-run with Unl as it ships read 67 of 90, with fewer decisions crossed and 19% more tokens.
Can Unl's context override my instruction?
It is not built to. Unl's decisions sit beside your message as labelled background, each with its reason. What you asked for stays the instruction; the decisions tell the model what you already settled while it does it.
What this is
Think inside your AI world — you stay in command
Save the thoughts, decisions and targets worth keeping, each with its reasoning, carried into every AI session the moment they matter. A new unit of exchange between you and your AI: the Settled Why with standing that travels. Unprompted.
Your whole AI world. What you decided at the epicentre. It plugs into Claude, Claude Code, ChatGPT and Cursor as an MCP connector — quick to connect, in a couple of steps.
MCP native·Human settled·Model agnostic·Your data
Reflections
Connect Unl to keep useful thoughts with their why.
Kept with its reasoning, without becoming a rule unless you settle it.
The full product, open. Free at launch.