Klaat Code

The deciding trait is what it does not charge you for. Tool rounds are unlimited and free, and only the things you type count against quota, which quietly removes the habit most people have developed of rationing questions and cramming three requests into one badly worded turn.

That changes how you work more than any feature on the page, because the cost of asking again drops to nothing. What you give up in exchange is control of everything behind the prompt. Pick it if the meter has been living in your head. Pick a key-based terminal agent if you would rather own the bill.

The problem is reproducibility. Model selection happens per message on the far side of the network, with automatic escalation when a task turns out to be harder than it looked, and nothing lets you pin a tier or read back which one answered. A run that worked yesterday is not a run you can repeat.

That matters most where the tool is aimed: if you are optimising cost, you cannot attribute a cost to a decision you neither made nor can inspect. What it does right is escalating at all. A router that never upgrades is cheaper and worse, and this one at least admits when a task got harder.

1. This is a real answer to the context problem rather than a larger window. Text search returns what matches; a call graph returns what participates, which is a different relation and the correct one for a change that has to compile. 2. Planning the read order before reading anything is the part that stands out.

2. No measurement accompanies it. There is no published comparison against a retrieval baseline on the same tasks, which is precisely the experiment this design asks for and would not be difficult to run. The architecture is well chosen and entirely unevaluated, which is a common pairing.

Charging for messages rather than tokens is a genuinely good idea, because it sells a unit the customer understands and can predict. What undermines it is that the unit has no number attached to it. A buyer cannot compare this against anything, so the sale needs a conversation, and conversations do not scale to individual developers.

Moat: the routing data, if usage ever accumulates enough of it to make the classifier better than a competitor's. Likely acquirer is an inference provider that wants a demand aggregator. Position: wait for a rate card, then evaluate the product instead of the promise.

Somebody on that team has been shouted at by a finance department. Warnings while a session runs, a ceiling per phase, a hard stop on the session, and a cost limit on an unattended run are four controls I usually have to invent myself and enforce with a wrapper script nobody maintains.

What is absent is everything above the individual. No single sign-on, no provisioning and no per-team reporting, so sixty seats means sixty accounts I did not create and cannot see across. Not yet: the spend controls are excellent, and I need one place to read them from before this goes past a pilot.

This is the shape I distrust on principle. An open licence on a terminal client is not ownership when the router, the index and the pricing all live on a server I cannot run, and the source I am permitted to read is the part that does the least work in the whole system.

I cannot supply a key, nothing runs on my own hardware, and there is no protocol for attaching the servers I already operate. Grudging respect for the honesty of the README, which says plainly that the client is a thin terminal to a service. It is exactly that.