Docs
kelve.aiGitHub

kelve and your agent

kelve has no model of its own. Anything that needs a model becomes a job for your agent. Your agent does it with its own model, on your account. kelve checks every answer before it shows you anything.

How a piece of work goes

  1. You ask your agent to show you something. It opens it in kelve, in a new browser tab.
  2. kelve reads it without a model. It walks the files as git does, masks secrets, and finds each file’s parts by the shape of the code.
  3. kelve makes jobs for what needs a model. The first job is the way in, the order to read it. Then come jobs for:
    • Which ideas a file uses.
    • A file or folder read plainly.
    • An entry the shared graph doesn’t have.
    • A note you left.
    • A lesson your agent learned.
  4. Your agent takes each job, does it, and hands in its answer.
  5. kelve checks the answer’s shape at once. kelve sets right what it can itself. An answer that still doesn’t fit goes back to your agent, with what’s wrong. If it’s wrong again, kelve puts it aside.
  6. You see it as soon as it’s taken. An entry goes to the shared graph, which checks each claim against its source before anyone else reads it.

Your agent keeps going until there are no jobs left. With a dozen files or more, your agent shares them among subagents if it can, so you see the work sooner.

What comes first

What you said to do next comes first. Then comes everything for what you opened last, in this order:

  • A note you left.
  • Its overview.
  • Its files one by one, each read plainly before the next.
  • Its folders’ pages.

Then comes what you asked for in other things. Reading for things you opened before waits until you open them again. See Your agent’s list.

Several agents at once

open gives your agent the number of what it opened. Your agent then takes only that thing’s jobs, by that number. So agents in different tabs each do their own. Copy what to say names a tab’s thing by its number too.

What a job gives your agent

Each job comes with its instructions, the material to work on, and the shape its answer must fit. Text from your files or the web is marked as material, never instructions. So nothing in what you open can tell your agent what to do.

The shared graph serves the instructions for each kind of job, by version. So every kelve asks for the same thing the same way. Each version, and each answer’s shape, comes in full once a conversation, then by name alone.

What it costs you

kelve’s work runs on your agent’s model, on your plan or your key. Most ideas already have an entry in the shared graph, so your agent writes only what the graph lacks. kelve reads only what you’ll read. It reads the way in first, then each part when you ask. What one person’s agent reads of a project can be kept with it, so others don’t read it again.

What kelve never does

  • Run anything from what you open: no builds, hooks or scripts.
  • Start your agent, or keep it going after it stops.
  • Change anything outside the project, except how your agent reaches it. That’s kelve in your agent’s MCP settings, and one hook there, each only with your yes.