Skip to main content
A machine is a computer that can run agent work — your own computer through the desktop app, or a server your workspace enrolls. Settings → Computers is where you see whether a machine is ready, control the Engine that runs work on it, and check which machines your workspace can reach. This page is the home for all of that; environments (the folder a project runs in) are set from each project, and server enrollment has its own page.
Most of Settings → Computers manages the computer you are sitting at, so it needs the Cogno desktop app. Opened in a browser, only the Your machines roster renders. Open the page from the desktop app to see This machine, Startup & background, and server enrollment.

What’s on the Computers page

On the enrollment approval page, select Back to app to leave code entry, review, or the completion screen and return to Cogno. Leaving does not approve a pending machine or undo an approval you have already submitted. The public /demo uses sample machines. Exploring it does not re-register your computer, remove its Engine credential, or reconnect your saved servers. Returning to your workspace resumes the usual machine registration checks.

Sign in to Claude Code before creating a Task

Open Settings → Computers in the Cogno desktop app to sign in on this computer. You do not need to create a Task first.
1

Choose an agent

Under Agent connections & accounts, choose Claude Code or Codex.
2

Open sign-in

In the Sign-in row, select Sign in. The button is available once this computer’s agent executable has been read.
3

Finish in the terminal

Complete the provider’s login in the terminal that opens. Opening the terminal alone does not mean you are signed in. If the terminal cannot open, the error appears in the agent’s settings so you can try again.
Cogno uses its bundled CLI for sign-in when it can run on this computer. Otherwise, it uses the CLI selected in this computer’s agent settings, including a custom path. Older Engines without executable settings use an installed CLI. This signs in on the computer running the desktop app. The browser does not show the local sign-in row. For a Task that runs on another computer, sign in on that computer instead. See Run agents on a task.

Choose the preferred agent for background runs

Open Settings → Computers in the desktop app. The Background runs summary at the top shows your preferred agent, or says no preference is selected. Select Change agent to reach Preferred agent for background runs, then choose Claude Code or Codex. Saved confirms the change was saved. This preference applies when a computer has both agents. A computer with only one uses that agent. Tasks you start yourself use the composer. The tabs under Agent connections & accounts manage sign-in and connections independently. In a browser, open the desktop app to change this preference. Desktop sign-out notifications check your default coding agent. For Claude Code, the check and its Sign in action use the default Claude account on this computer. A failed or unavailable authentication check does not confirm a sign-out and does not trigger that notification. Use the agent’s Sign-in row to inspect or repair its connection.

Choose models for background runs

In the desktop app, open Settings → Computers → Background runs. Choose a model separately for Claude Code and Codex. These saved settings apply to runs Cogno starts; Tasks you start yourself use the composer. Choose Optimized · Recommended to select a model and reasoning effort for the work from the executing computer’s available models. Automatic lets Optimized choose the effort. Selecting an effort limits the candidates to models that support it. Among the candidates judged able to finish the work at the quality it needs, Optimized picks the one expected to use the least, counting a likely retry, and learns from how earlier Optimized runs in the workspace went. The conversation shows the selected model, effort when present, and the run’s completion or failure together. Optimized supports background runs on a registered computer with a signed-in Claude Code or Codex account and a supporting Engine. Cogno-credit and managed cloud runs do not support Optimized. If the model list or selection service cannot be read, the run reports an error. Select a specific model to use a manual choice. Existing conversations keep their model.
Model selection sends a bounded work description and the available model options to Cogno’s selection service through a zero-data-retention AI route. It does not read repository files, credential files, or conversation history. Cogno records selection outcomes, candidate probabilities, latency and reported usage for reliability monitoring. These records exclude the work description and raw provider responses.
You can also ask Cogno to start a Task with a specific model and effort, or with Optimized. That request applies to the Task Run without changing your saved background-run defaults.

Control execution access

Each entry under Your machines has two access settings: Choose Ask for approval, Auto in workspace, or Full access for each setting. The user-initiated setting is a machine maximum: each Session also has a Session access choice in the composer, and the Engine uses whichever choice is more restrictive. The autonomous setting directly controls unattended queue turns.
Full access can run commands and change files outside the Session workspace without native approval. It may also reach files and credentials available to the Engine’s OS account. Cogno asks you to confirm before enabling it, especially for autonomous runs that can start while no one is present.
An older Engine keeps the safe Auto in workspace boundary and does not offer Full access; update the Engine before using Session access controls.

Keep the Engine running

The Engine is the always-on local runtime that runs agent work on a machine. Under This machine, its state reads Running, Stopped, or Not running, with Start, Stop, and Restart controls.
Closing the window or choosing Quit Cogno leaves the Engine running. To stop agent work on this machine, use This machine → Stop. Turn on Startup & background → Open Cogno at login to open Cogno when you sign in.
On macOS, the Engine automatically prevents idle sleep while agents are working, including background agents. No settings change or additional permission is needed. Normal idle sleep resumes when all agent work has stopped; waiting for a future scheduled run alone does not keep the machine awake. The display can still turn off and the screen can lock. Closing the lid or manually putting the machine to sleep can still interrupt work. A machine that stops reporting for two minutes counts as offline, and runs routed to it wait in the queue until it is back.

Update while Tasks are running

Open Settings → Preferences → Desktop updates in the desktop app to choose how this device updates. These settings do not change other devices. Available updates download in the background even when automatic installation is off. Downloading keeps the app and your work running. Confirmation happens before installation, when the update needs to stop work and restart the app. Temporary connection failures are retried automatically up to twice during update checks and downloads. If all attempts fail, the update control shows the error. Signature verification and permission errors require action and are not retried.
Turning off Confirm before interrupting work allows updates to stop recordings, running Tasks, and terminals without asking. Save your work before disabling it. A failed or unconfirmed shutdown still prevents installation.
When confirmation is enabled and Tasks are running, choose Update later to keep working, or Interrupt and update to continue the update. After restart, an update-interrupted Task can continue once in the same Session. Waiting or new human messages take priority and replace the automatic continuation. A manually stopped or cancelled run does not restart automatically. A closed Session or changed folder requires manual review instead of starting a new Session automatically. The Engine being replaced must already support recording the update interruption. Installing that support for the first time cannot recover an interruption reported by an older Engine; continue that Session manually if needed. Remote Engine updates keep waiting for idle work.

Read the Readiness line

This machine shows a Readiness line. When it reads Ready, the machine can pick up work. When it does not, the details list exactly what blocks it — for example:
  • neither Claude Code nor Codex available, or a provider signed out — with claude auth login or codex login shown as the fix
  • git or Node.js missing
  • the GitHub CLI signed out
  • the disk almost full
Fix what the line names, then the machine returns to Ready on the next check-in. A new run uses a healthy supported CLI and does not prefer one provider when both work. Reset & re-register applies to the Cogno machine credential; for a coding-agent sign-out, run the provider-specific login command shown in the readiness details.

Environments live in the project

An environment is a folder on a machine plus the settings work runs under there (AI tool, model). Settings → Computers shows which folders this machine has registered, but you do not add a project’s folder here — each member connects their own folder from the project’s settings. See Set up a project for connecting a folder or cloning repositories.

Run a machine on a server

To run work on a machine you are not sitting at — a Linux box, a cloud VM — enroll it as a server. Set up a new server walks the ssh enrollment, and cogno doctor on the server reports each readiness check with the command that fixes it. See Run Cogno on your own server for the full flow, and Enterprise remote execution for fleet setups.

Reference

Routing

How a run chooses the machine and directory it lands on: A task that has already run keeps its environment, so later runs land on the same machine and directory. When many requests arrive at once, Cogno counts work already placed on a machine but not yet started as load, so they spread across your machines instead of queuing on one. For the states a run itself moves through (pending → claimed → running → completed), see Run agents on a task.

Engine states

Machine status

Troubleshooting

For anything not listed here, see Troubleshooting.