Getting started
A place for your assistant.
Comis runs on a machine you control. Here’s what to prepare, and what is still ahead.
There is no published release or supported public installer yet. This guide describes the target setup. Verified download links will appear here when a release is available.
Choose your host.
The target baseline is 4 CPUs, 8 GB of memory and 20 GB of disk for the chief of staff and one additional agent. A local model needs its own resources on top of that.
Linux
The target host runs Docker on x64 or arm64. Linux uses gVisor for additional workbench isolation. The installer is designed to check the host, explain its plan and ask before making changes.
macOS
The target uses Docker Desktop, OrbStack or Colima. It runs the services in Linux containers, without the Linux host’s gVisor layer. Native macOS installation qualification is still pending.
Windows
The target is Windows 10 22H2 or Windows 11 through WSL 2 and Docker Desktop. The planned PowerShell installer prepares that environment. Native Windows installation qualification is still pending.
Bring a model and a channel.
- Model access: an API key for Anthropic, OpenAI or OpenRouter, or a local model. Model providers can charge for usage.
- A conversation: a Telegram bot token for the default channel.
- Your limits: choose budgets and review the rules for actions that need approval.
The planned setup.
- Review the installer. The release will include a script and checksum. Read the script and its plan before running it.
- Connect your model and channel. The Gate holds credentials outside the agent’s workbench.
- Say hello. Your chief of staff receives the conversation and coordinates workers.
Download commands are intentionally absent until they point to a verified release. The source repository is currently private; this website does not require GitHub access.
Know the boundaries.
Self-hosting gives you control of the machine and files. It also makes you responsible for access, updates and backups. Hosted models and connected services still receive the requests you choose to send. Read the security model before connecting accounts.