Comis is an open-source platform designed for deploying self-hosted AI agent teams directly into existing messaging channels. It supports a wide array of communication platforms, including Telegram, Discord, iMessage, IRC, Line, Signal, Slack, and WhatsApp. The platform is built with a strong emphasis on security, incorporating kernel-enforced execution sandboxes (using bubblewrap/sandbox-exec), input/output scanning, trust-partitioned memory, and per-agent tool restrictions. Additionally, Comis ensures supply chain integrity through sigstore attested releases and bundled dependencies, preventing unexpected runtime issues.
One of Comis's notable features is its cost-saving mechanism, which includes a dual prompt architecture and active cache fences, leading to significant prompt cache savings. The project claims an 81% prompt cache saving and a 94% cache hit rate on warm turns. For context management, Comis utilizes a 10-layer context engine that evicts stale content and compacts older history into structured summaries, ensuring that critical references are re-hydrated without silently dropping messages.
Comis is implemented in TypeScript and is designed for production use on Linux hosts, with Docker support available for both development and deployment. The latest release is v1.0.42, which includes security fixes and CI improvements.