What do I need to start?
A server or a computer, access to a model, a Telegram bot. Root is not required: the same command installs into your home folder.
How much does it cost?
Installing it for yourself or inside your own company is free. You pay your model provider: your own subscription or key. Hosting, setup help and support are sold separately. A CapyAgent-run model subscription is in development and cannot be selected yet.
Where is my data?
On your machine: history, notes, tasks, settings. Only what you connect leaves it: the request to the model and any service whose keys you added. No telemetry. Voice notes are transcribed on your machine, audio is not sent anywhere.
Which models can I use?
A ChatGPT subscription (sign in with a code from your phone), any OpenAI-compatible endpoint (OpenRouter, DeepSeek, Qwen, Kimi and others), an Anthropic API key, Gemini, plus a model on your own machine through Ollama or LM Studio. A Claude subscription cannot be used in place of an API key.
Do I need a VPN?
The wizard checks Telegram and your model with a real request. If one of them does not open from your server, it offers a proxy, or our VPN+ subscription, which carries only Telegram and the model, not the whole machine.
Can I run it without a server?
Yes. Install the desktop app on Windows, macOS or Linux, or run the same command on your own computer. One catch: while the computer is off or the app is closed, the agent is silent and schedules do not fire.
How do I update?
On your say-so: the agent does not look for updates on its own. Run sudo CAPYAGENT_UPDATE=1 /opt/capyagent/bin/capyagent update. It keeps the previous version and your data first and rolls back on failure.
Is it safe?
Anything that gives access to the machine (commands, files, screen capture) is off after install. A risky action runs only after a one-time "yes" that names the action. Destructive ones (wiping the disk root, a fork bomb) are refused in any mode. Strangers cannot message the bot. The limits live in code, not in a request to the model.
What does it not do?
It does not listen to your microphone. It does not watch your habits to make suggestions. It does not pay for you or use your bank. It writes first only to chats you named in advance, and no more often than the limit you set.
Can two people or a whole team use one agent?
Yes, add them to the allow-list. Each gets their own conversations, while memory and access are shared in one agent. If they must not be shared, run a separate instance.