cerveau docs site github
Start here

Quick start

One session from empty workspace to committed change, and a tour of the panel you will spend your time in. Assumes you have installed and have a model serving.

Your first session

Start the core and open the panel:

crv
# → http://localhost:7700

Pick a workspace first — the folder icon in the status bar, or:

curl -X POST localhost:7700/api/config/workspace \
  -H 'content-type: application/json' \
  -d '{"path":"/home/you/projects/thing"}'

Then, in Discussion mode, describe what you want. Discussion is for deciding, so keep it short and let the model ask. When the shape is agreed, it calls commit_plan and a plan strip appears above the composer.

Switch to Autopilot and let it run. The committed plan is injected as guidance; it follows the intent and adapts. Watch the tool cards land — each one shows the real command and its real output.

The same thing from a terminal

cd ~/projects/thing
crvcli ask "add tests for the parser"
crvcli ask -mode brainstorming "how do other projects handle this"

The CLI is independent of the panel: it uses the folder you are standing in, and the panel's picker never redirects a session started there.

The panel

Status bar

The brand, a dot-matrix version readout, and on the right the hardware chip and the status orb. Hover the orb for the live stack: every component with its URL and state, the context budget, and GPU/VRAM/CPU/RAM meters. A red orb means something is down, and the panel says which.

Mode dial

Discussion, Brainstorming, Autopilot. Changing it changes which tools exist for the next turn — see The three modes.

Plan strip

Appears once a plan is committed. Steps tick green as they complete, the counter tracks done against pending, and a failed step shows as failed rather than disappearing. When every step is done the strip celebrates and collapses.

Tool cards

Live cards show the real command and its real output as it runs. Errors surface as actionable cards with the reason expanded — never a silent spinner, never a log wall. A guard denial appears here too, with its tier and hint.

Memory view

A graph of what has been stored, with provenance: click a fact to see the event it came from and what it superseded. Promotion candidates awaiting review are listed separately — see Memory.

Composer

The message bar, with a sampling dial and an attachment knob. ask_user questions appear as cards above it and block the run until answered.

Habits that pay off

Plan in Discussion
A committed plan gives Autopilot a spine and gives you something to check before it starts writing.
Let it ask
ask_user exists so a wrong assumption costs a minute instead of an afternoon. Answering promptly is the highest-leverage thing you do.
Watch the zone
If the window report is amber every turn, your model_ctx and your engine probably disagree. See Configuration.
Use instant sessions for questions
They never promote to long-term memory, so a throwaway question does not colour the store.
Check the guard, not the model
If something did not run, the tool card says which rule stopped it and why. That is faster than asking the model what happened.

Next