AI agents
Connections, models and limits
Connections are the accounts and models your agents run on: a subscription you sign in to through its official tool, an API key, or a model on your computer. This page also covers how OmniGet shows the usage limits of those accounts (the limits strip and the usage icon) and the execution settings for stuck agents and long contexts.
Where it is
- LLM › Configure › Connections (the page is titled AI connections).
- LLM › Configure › Models and limits.
- Settings › General for the Limits strip and the Usage icon in the menu bar.
- The Usage monitor button at the top right of every LLM tab.
Add a connection
- Open LLM › Configure › Connections and start Connect my AI.
- Choose how to connect:
- Use my subscription: Claude Code or Codex. You sign in through the official tool in a visible terminal. A chat subscription is not API credit.
- API key: OpenAI, Anthropic, OpenRouter, DeepSeek, Google Gemini, Groq, xAI (Grok), Mistral, SiliconFlow or Requesty. The key is saved in the system vault.
- Local model: Ollama, LM Studio or llama.cpp.
- Follow the steps (Choose, Connect, Prepare agent, Test and open). The test sends one real short message and may use a little quota.
Models and limits also has Sign in with OpenRouter, which connects an OpenRouter account in the browser without copying a key.
Several accounts of the same tool
Under Advanced: CLI accounts, fallback and limits, New account creates an extra Claude Code or Codex account. It creates an empty configuration folder; click Enter and sign in inside the tool itself. Each account then appears as its own driver in Threads.
The same section has a Fallback chain and Automatic rotation between your accounts, which is off by default. As the app itself warns, Anthropic’s published limits assume ordinary, individual use; whether rotating between subscriptions you own fits a provider’s terms is your call. OmniGet does not raise or change any provider’s limits.
See your limits
Usage on the Connections page
The Usage section reads the session logs that Claude Code and Codex keep on your computer and shows each account’s five-hour and seven-day windows. A value is reported when the tool wrote it in its own log, and estimated when OmniGet computed it.
Limits strip
A small always-on-top strip on a screen edge, with one ring per coding assistant: how much of each usage limit is used, when it resets, and whether the assistant is working, waiting or done.
- Open Settings › General › Limits strip (or click Usage monitor in LLM).
- Turn on Show the limits strip, then, under What it may read, turn on each assistant you want.
| Option | Default |
|---|---|
| Show the limits strip | Off |
| Screen edge | Top (you can also drag the strip to another edge) |
| Size | Medium |
| Show percentage | On |
| Show pace | On |
| High contrast | Off |
| Warn when on pace to run out | On |
Nothing is read and no window exists until you turn it on. To read a limit, OmniGet uses the login the tool already keeps on your computer, read-only, and asks that provider’s own server, at most every 5 minutes. Claude Code, Ollama and LM Studio are checked against the real tools; Codex, Cursor, Copilot, Grok, Kimi and OpenCode are marked beta (written from their documentation).
Usage icon in the menu bar
An icon in the menu bar (the system tray on Windows and Linux) that shows the usage of your Claude Code and Codex accounts. It is on by default on macOS and off on Windows and Linux. Turn it on or off in Settings › General or from the tray menu.
| Option | What it does | Default |
|---|---|---|
| Icon shows | The session window, Weekly or Most used account. | The session window |
| Show percentage | The number next to the icon. | On |
| Loop in colour | A coloured icon. | Off |
| Alert thresholds | When the icon turns orange and red. | orange at 80%, red at 95% |
Click the icon to open the usage panel. Nothing is read until you pick an account.
Models and limits
LLM › Configure › Models and limits lists the providers and their models, with Manage AI connections. Under Advanced settings:
Context pruning
Between turns, old tool outputs that no longer bear on the task are replaced by a marker, so long conversations use fewer tokens.
| Option | What it does | Default |
|---|---|---|
| Context pruning | Turns pruning on. | Off |
| Judge | Local runs a small model on your computer and sends nothing out. Jev (TypeSafe) is a remote service that needs a TypeSafe API key. | Local |
| Start judging at | The context size, in tokens, before pruning starts. | 50,000 |
Stuck agents and approvals (watchdog)
When a silent agent is stopped, how long a question waits for you, and when you get notified.
| Option | Default |
|---|---|
| Stop after silence (minutes) | 10 |
| …while a tool runs (minutes) | 30 |
| Start timeout (seconds) | 90 |
| Job time limit (minutes) | 60 |
| Question wait in chat (minutes) | 30 |
| Question wait in jobs (minutes) | 0 (waits until the job’s time limit) |
| Notify after (seconds) | 5 |
| Resume by itself after a restart | Off |
Common problems
The limits strip shows nothing for an assistant
Turn on that assistant in the strip’s settings, and make sure you are signed in to the tool itself. Beta readers can be wrong; the provider’s own page is the reference.
The usage numbers differ from the provider’s page
Values marked estimated are computed from local logs and public prices; plans and discounts can make the real bill different.