knowop🔌

One command to log in to KnowOp, connect your computer, and run your agents and GPU jobs where Chat can reach them.

Quick Start

# Mac:
#   # Required for quick install.
#   brew install uv

# Ubuntu/Debian:
#   # Required for quick install.
#   sudo apt-get install -y curl
#   curl -LsSf https://astral.sh/uv/install.sh | sh

uvx knowop connect

uvx knowop connect runs knowop without installing it. To keep the command for every day, install it:

uv tool install knowop
# Alternatively: python -m pip install knowop

Use

knowop connect          # this computer: log in, enroll, serve its terminal and commands
knowop run claude       # Claude Code in this folder, recorded, reachable from Chat
knowop run codex        # the same for Codex
knowop job submit -- python train.py   # a job on a GPU node, from this repository
knowop job submit --machine lab-box -- python train.py   # on your computer lab-box
knowop status           # server, account, your machines and the runs live on each
knowop disconnect       # stop this computer's runs, then disconnect it on the server

A run reaches Chat with your account's key, not through knowop connect: it keeps running after Ctrl-C ends the connect, which then names the runs it leaves. The web app shows a machine connected while one of its runs is live. knowop run stop --all stops the runs on this computer; knowop disconnect also revokes its connect and stops every run of yours that names it, wherever it runs.

A job can also run on a computer of your own: start it with knowop connect --jobs there, push your branch, and knowop job submit --machine NAME -- CMD from the repository. That computer fetches the pushed commit into a fresh worktree and runs CMD under a memory cap (a systemd-run --user scope), only for your account; its log and results come back through the server, which keeps none of them.

A job runs with the full authority of the Unix user running knowop connect --jobs, including access to that user's files, credentials and inherited environment. The worktree and memory cap do not restrict that authority. GPU selection through CUDA_VISIBLE_DEVICES is cooperative: submitted code can change it. Run untrusted code under a separate Unix identity or external isolation that enforces the access restrictions you need.

knowop connect prints a code; approve it in your browser, on any device. That one approval logs you in, enrolls this computer and confirms it. Until you press Ctrl-C, the computer's terminal opens in the web app, and it runs the commands the server's assistant sends; --no-exec serves the terminal only. knowop connect acme does the same on https://acme.knowop.ai. A computer with no browser, such as a box, pipes an enrollment token to knowop connect instead, and joins with no login.

Plain knowop login logs in to https://alpha.knowop.ai; knowop login acme to https://acme.knowop.ai, and knowop login URL to any server. Use it on a computer you will not connect, where only your agents work. A later knowop connect there enrolls the computer with that login's key, and it waits until you confirm it in the web app. A first knowop run logs in too.

Commands

Every command takes --help; knowop run claude --help and knowop job submit --help show the runner's own.

knowop keeps one login and uses it for every command. It saves the key as a trax profile, and trax's own current profile stays as it was: knowop never acts on a server it did not log in to.

Direct terminal link

Every install includes aiortc for a direct link between the browser and this computer. Terminals use the server relay when a direct link is unavailable.

Layout

knowop calls the server through trackinizer's Client, as trax does, and never runs trax. trax machine and trax job say they moved out of trax.

Development

uv sync --all-groups
uv run pytest

See also

Sibling projects in the rekursiv-ai family:

• trackinizer — Centralized agent database for tracking inquiries, work, and the evidence behind conclusions; knowop's trax and its client come from it.
• sagent — The self-mutating multi-provider coding-agent CLI and typed Python library.
• wesearch — Web search, resilient page fetch, and scholarly-paper lookup without a browser stack.