Chat

Chat is an assistant called Nemu that runs on the models you connected in the console. It searches the web, reads your files, queries your databases, drives a real machine, and hands work to subagents that run in parallel. The model list is entirely yours. There is no curated set. Whatever providers and models you enabled in the console are what the picker offers, so what chat can do is decided by what you connected.

​

It keeps running without you

Work does not happen in your browser. Sending a message enqueues a job, and the job runs on our side whether or not the tab is open. Close it, put the laptop to sleep, come back later. When you reopen the chat it reattaches to the running job and replays everything that happened while you were gone. Two honest caveats. The pointer to a running job is stored in that browser, so another device sees the answer once it is saved rather than watching it stream. And the live replay buffer is kept for one hour, after which the finished message is still there but the keystroke by keystroke replay is not. If the agent needs an answer from you, it waits 90 seconds and then suspends rather than holding a worker open. The message is saved as suspended and you pick it back up with /resume.

​

Modes

Five, from the composer. They change how long the agent may work, how many subagents it runs at once, and which tools it may touch. Normal, Plan and Fast run up to 300 steps and 45 minutes. Eco runs up to 40 steps. Max runs up to 600 steps and 90 minutes.

​

Fast

Fast scales its process to the request. A question, a fix or a single file gets answered or written directly, with no planning first. A large build with independent parts, such as a multi-file app with components, libraries and routes, is split: the agent writes a short shared contract (the file tree, the types and props every file agrees on, the data shapes and the naming), then hands 3 to 5 builders a disjoint set of files each and lets them write at the same time. On a Computer the agent scaffolds the project and installs every dependency before the builders start, and they write straight into it; without one, each file becomes its own artifact. Builders run at low or medium reasoning effort while the agent itself keeps the effort you picked. When they are done, the agent runs one check over the whole build, such as a typecheck or a build command, fixes what does not fit together, and reports.

​

Eco

Eco is the cheapest way to use the same model. Reasoning effort is held at low whatever you picked, replies are short, a single reply can write at most 8,192 tokens, and a turn stops after 40 steps. There are no subagents, no research agent, no plans or status updates, and the agent does not look at or click through the preview page, though it can still read the browser console. It can still build small things: an artifact, or a small project on a Computer with its files written and its commands run directly. For a request that needs more, it does a reduced version and says which mode suits the full one.

​

Checking the preview

In every mode, the agent’s look at a running Computer preview is bounded. It can take 12 preview actions in a turn, and after 3 failed in a row it stops checking and tells you what it could not confirm instead of trying more selectors.

​

What it can reach

​

Slash commands

With a Computer attached you also get /logs, /stop, /usage and /publish.

​

Context

The header shows how full the context window is, broken down by messages, tools, input, output and reasoning. At 90% the conversation is compacted automatically and older turns are archived. Nothing is lost: the agent can still search the archive when it needs something from earlier.

​

Deleting a chat

Deleting a chat deletes its messages, and if it is a Computer chat it deletes the project too, with every file in it. There is no undo and nothing is kept, so the confirmation says so before you agree to it. If you want the project without the chat, publish or download what you need first. See Computer.

​

When we are busy

Chat jobs queue. If the queue is deep you see your position. Plans from Pro up have a priority queue and bypass the load shedding that protects the platform when it is under pressure. Separately from the queue, we watch the load on the machine that runs the agent. If it stays busy for several minutes we stop starting new work for a while, and a message you send then waits with a note saying we are catching up. Anything already running is left alone and finishes normally, and queued messages start on their own as soon as there is room. Nothing is lost and nothing needs resending.