Run the clean brain loop.

Octopus lets the model ask for a cognitive Need, then lets an intelligent tentacle decide how to supply Feed. This page starts from the local product path.

Open local app Run locally

Local app path

1. Start Octopus

Open the local app after starting Octopus. The first screen shows the current Goal, Need, Feed, and pet state.

2. Change the Goal

The page only writes the Goal. Model and provider settings stay in the local runtime, not inside the Need text.

3. Watch Feed arrive

The clean brain emits a Need. A model-backed tentacle enters the environment, acts, checks itself, and returns compact Feed.

The product loop

GoalUser changes the objective.

NeedThe clean brain asks what it needs.

FeedA tentacle supplies evidence.

Local path

cargo install --git https://github.com/dangoZhang/Octopus octopus-core --locked --bin octopus --force
octopus start --open

The launcher prepares state, editable seed tentacles, the profile registry, three beats, the pixel Octopus, and the local app at http://127.0.0.1:8765/app.html.

octopus first-run "make this repo easier to use"
octopus chat "make setup clearer"
octopus preflight

The app page keeps user input to Goal. Need, route choice, provider routing, repair, and harness evolution remain agent work and show up as current Need, current Feed, pet color, and Output.

What to expect

Brain stays clean

The model context is Goal, memory, Needs, and Feeds. It does not carry shell syntax, tool routing, or install details.

Tentacles think

Each tentacle owns prompt, metadata, code, permission, traces, and an evolution surface that can be improved outside the kernel.

Evidence is visible

doctor, report, start --check, and preflight show readiness without turning those checks into the first user path.