Y
01
02
03
04
05
06
07

Control desk

One request.
One delivered project.

Ymir takes what you write, from a one-line fix to work that spans several coordinated batches, and carries it through by chaining Claude Code runs. For every step, it composes the agent that fits: the model, the effort, the skills, the servers that agent needs.

See how it decides No account, no sign-up

What the record holds

cost
tokens
duration
skills loaded
output
diff
failure
its cause

Recorded run by run

01 · The problem

A project is not a session

One agent session, you tune it by hand and you watch it with your own eyes. A project is dozens of runs. The one exploring an unfamiliar repository does not need the same model as the one fixing a red test. The one writing a migration does not need the same skills as the one reviewing a diff.

Today that tuning gets redone every time, or it does not happen at all. That gap is what Ymir fills.

02 · The orchestrator

Someone decides, and it is not you at every step

At the centre of Ymir sits the orchestrator. It reads the project memory, picks the smallest answer that will do, launches it, watches it while it runs and steers it back when it drifts. When nothing off the shelf fits, it builds the recipe or the workflow to measure.

01It reads the project memory
02It sizes the answer
03It composes the agent for the step
04It launches and it watches
05It steers back when things drift

That is what separates Ymir from a well-equipped run launcher.

03 · The proportionate answer

Five sizes, and always the smallest that answers

Firing off a whole workflow for a typo is the worst failing this kind of tool can have, and it is the one it leans towards on its own. Ymir is built against that slope. Every request gets a size, and going up one notch is justified in writing.

When a request goes past what one mission can carry, Ymir suspends and names the reason instead of doing half of it without saying so. You split it up, or you launch the workflows one at a time.

A tool that admits what it cannot do costs you less than a tool that tries anyway.

C0
Answering is enough. No run, nothing written.
C1
One agent, one pass.
C2
A short chain, a few steps.
C3
A whole workflow.
C4
Several coordinated workflows. Ymir says so and hands back to you.

04 · Two opposite users

The same product for two people who have nothing in common

Two settings, and the first one has four rungs. You can move them mid-course.

Creativity, four rungs

0
executionStrictly the request.
1
proposalsPlus a written list of what it would have done.
2
initiativedefaultPlus the reversible adjacent.
3
directionThe right to settle the breakdown.

Mode, two values

copilotIt stops on every structuring decision and hands back.
autonomousIt decides and records. It only stops on a hard refusal.

You know exactly what you want.

You keep the wheel, the agent applies.

executioncopilot

You have never written code.

You let Ymir propose the way to do it.

directionautonomous

05 · The guardrails

Autonomous does not mean without guardrails

Merging is never automatic.

In neither mode. You are the one who asks for it.

Every run has its own copy of the repository.

On its own branch. It commits, it does not push, and that is not an instruction handed to the model: the working copy simply has no valid push address.

06 · The record

What every run cost, and what it did

Without those records, saying a recipe has drifted stays an impression. With them, it can be shown. The detail is at the top of the page.

07 · The catalogue

You never start from a blank page

Ymir knows 203 skills, gathered from 8 sources you choose. It carries 22 agent recipes and 10 workflows ready to use, and everything a recipe can set is set from the interface.

203skills
8sources
22recipes
10workflows
A project factory, built on Claude Code. Open the product