Bring your own key
Connect your own Anthropic or OpenAI API key to unlock Claude and GPT on any plan, billed by your provider instead of Clusy.
If you already have an Anthropic or OpenAI API key, you can hand it to Clusy and run the agent on it. The models you unlock this way sit in the picker next to the rest, but their inference bills your provider account, not Clusy, and consumes no Clusy model allowance.
This also means they ignore your plan. Bring your own key and you can run Claude Opus on the free tier. That's the honest answer to "can I use Claude without upgrading": yes, if you pay Anthropic for it directly.
Every model call goes to your key — including the work you never see. Clusy's model allowance stays untouched.Connecting a key
Go to Settings → Integrations and paste your key into the Anthropic or OpenAI card. That's the whole setup. You can disconnect it from the same place.
What each key unlocks:
| Provider | Models it adds |
|---|---|
| Anthropic | Sonnet 5, Opus 4.8, and Fable 5.1, on your Anthropic API key. |
| OpenAI | GPT-5.6 Sol, on your OpenAI API key. |
Clusy also offers GPT-5.6 Sol and GPT-6 Sol as hosted models on the plans that include GPT. Connecting your own OpenAI key is the way to use the same model on any plan, billed directly by OpenAI instead of drawing from your Clusy allowance.
Connecting a key also switches your default model to that provider's flagship: an Anthropic key moves you to Sonnet 5, an OpenAI key to GPT-5.6 Sol, along with that model's default reasoning effort. Sessions where you'd already picked a specific model keep it. If you later disconnect the key, any selection pointing at those models drops back to your plan's default, so nothing is left pointing at a model you can no longer run.
What it costs
Clusy does not meter the BYOK model inference, on any plan. Your provider bills those tokens at its own rates.
The rest of the platform still bills the way it always does. Cells run in a Clusy sandbox, so GPU time is metered as usual, the agent's metered web tools still count, and on the free tier even CPU sandbox time draws down the allowance. A key covers the model, and only the model.
Your key runs everything the turn does
A turn is more than one model call. When it's running on your key, the calls the agent makes along the way — the chat itself, its self-verification, the sub-agents it spawns, what it extracts to remember later — all resolve to the same key. Clusy doesn't quietly hand part of the work to a house model. The flip side: everything the turn does shows up on your provider's bill.
What this changes about the picker
- Plan gating stops applying to these models. They're gated on the connected key instead. That includes the free tier, which is otherwise locked to Auto.
- They're normal picker rows. Same effort control, same per-message switching, marked with a Your key badge so you can see which lane a message is about to go down.
- BYOK Sonnet 5 has a slightly different effort surface. Hosted Sonnet offers Low, Medium, High, and Max; on your own Anthropic key it also offers Extra High (Anthropic's
xhighlevel). The picker always shows exactly which levels a model supports. - Your key is used for your work and is never logged by the agent. Clusy stores the credential encrypted and never persists or logs the plaintext key.
The other way to do this
If what you have is a Claude or ChatGPT subscription rather than an API key, the equivalent route is the local runner: pair your own machine and Clusy drives the claude or codex CLI you already have installed. The model work uses no Clusy allowance; sandbox and tool usage still follow the normal rules.