User-owned local path

Your machine.
Your model.

The Local Companion binds to loopback, authenticates the browser with a pairing token, and routes compatible local text inference from any entitled box. GoldCap alone prepares approved training data and starts user-configured compatible training.

Cracked-gray Curious Obmil carries a glowing device toward a smaller, sorrowful GoldCap with one broad main cap, exactly two smaller caps, and living gold veins in a device-local workshop glade
Curious brings the work home. GoldCap keeps the training path on the user’s own ground.
CURRENT RELEASE

Version 1.0.0

Released 2026-07-23 · 9.2 KB · Python 3.11+

SHA-256 60ae0416c913a725b54be8e33ce0fc2d2867c62150b86dbb431ae52b6827d3f9
Download decacap-local-companion.zip
127.0.0.1Companion binds to device loopback
Browser onlyEndpoint + token stay in local browser storage
ExplicitPreparation, training, and activation remain separate actions
WINDOWS

Install + pair

  1. Verify the SHA-256 above.
  2. Extract the ZIP to a user-owned folder.
  3. Install Python 3.11 or newer and a local runtime.
  4. Run start-windows.bat.
  5. Enter the printed loopback endpoint and token in Workbench Local Settings.
  6. Run health, model-list, and short inference tests.
Detailed Windows guide
MACOS + LINUX

Install + pair

  1. Verify the SHA-256 above.
  2. Extract the ZIP to a user-owned folder.
  3. Install Python 3.11 or newer and a compatible runtime.
  4. Run python3 decacap_companion.py.
  5. Pair from Workbench Local Settings.
  6. Run health, model-list, and short inference tests.
Detailed macOS/Linux guide
RELEASE NOTES

What this build contains

01

Removed fixed inference timeout

02

Per-user loopback pairing

03

Ollama and compatible local text-runtime support

04

Explicit local data preparation and training lifecycle controls

UPGRADE
  • Stop the current companion process.
  • Back up decacap-companion.json and the local data folder.
  • Verify and extract the new ZIP into a separate folder.
  • Carry forward only reviewed local configuration.
  • Restart, re-pair if the token changed, and rerun health/inference.
UNINSTALL
  • Stop the companion process.
  • Remove the extracted program files.
  • Remove the device-local Workbench endpoint/token from Local Settings.
  • Keep or delete local datasets, models, adapters, and logs according to your own retention decision.
  • Uninstall the separate model runtime only if no other application uses it.