Tools should become local nervous systems.

Most agents make one model carry intent, tool syntax, route choice, execution detail, and evaluation. Octopus keeps the main brain focused on Need, then lets independent tentacles supply compact Feed.

Install and run locally Read the tutorial Copy recipes

The Bet

Biological octopuses combine high-level intent, local arm control, and feedback. Octopus maps that pressure to agents: the brain asks for verification, reproduction, comparison, memory, forgetting, recall, or execution; tentacles choose how to satisfy the request.

What Changes

Clean brain

The main model sees Goal, memory, Need, and Feed. It does not carry command syntax, API trivia, or tool routing burden.

Thinking tentacles

A tentacle reasons near the tool, acts in the environment, and returns only the useful Feed to the brain.

Feedback supply

Feed traces, check history, and repair outcomes teach the harness which tentacle can supply better evidence next time.

The Loop

GoalThe user steers the objective.
->
NeedThe brain asks for cognitive supply.
->
FeedA tentacle returns compact evidence.

Runs Today

Local product

octopus start --open prepares state, seed tentacles, three beats, the pixel pet, and the one-page app.

Release evidence

start --check, provider matrix, benchmark evidence, and preflight records keep each release honest.

Why The Page Is Short

The product surface should feel like adjusting a Goal. Provider setup, tool selection, repair, preflight, and harness evolution remain visible as evidence, not as the first thing a user has to drive.