Silta: family assistant on Matrix with continuity across context limits

Silta is a self-hosted assistant serving multiple users from a single Matrix account, used by me, family, and close friends for the last three weeks. It runs on Claude Code with each person using their own Claude subscription.Silta answers everyday questions, conducts web research (presenting the results as PDFs), performs scheduled tasks to monitor/research something and notify you if needed.It uses a deliberate handoff & compaction procedure, writing the compaction summary itself with a custom prompt, so the thread of the conversation and its picture of the user is preserved across context limits. As it turned out, the assistant accessible in a continuous chat (contrary to ChatGPT / Claude Web sessions) builds up better context about you and your tasks, and provides more relevant answers and help.It runs on Claude Code as the only allowed way to use a personal Claude subscription. Using Claude Code also spares me developing a harness, at the cost of tracking changes as new versions of Claude Code are released. The assistant itself (the daemon, the channel plugin, the supervisor, and the deployment) is open-source, while the model and the harness are Anthropic's and proprietary.Silta is deployed to a dedicated VM with separate Linux users per session and systemd hardening, protecting sessions from each other. Claude Code's bubblewrap sandbox is used to protect the Claude credential and the harness's state from the agent. This does not protect from prompt injections coming from search results and potentially leaking user data.Matrix E2EE protects the conversation from a Matrix server operator, but everything in the context is of course sent to the model provider. Private content is never logged, and the operator is expected to respect the privacy and not read the session transcripts on disk. Additionally, it is recommended to use full-disk encryption and tell the users how their data is handled.Silta only works in its workspace with files explicitly shared by users in the chat. Giving it access to a user's own computer or services like email is not planned.Since Opus 5.5 has been released, it is possible to run a session on a $20 Claude Pro subscription. Before that, only Fable 5.1 was pleasant enough to talk to. When run on Anthropic API, a moderately used session extrapolates to $100-$150 per month, with each research run costing $6-$12.Silta is built with Claude Code and Silta itself. The design, review, deployment, and testing are mine.