Retrieval is keyword, relationship-graph and metadata driven, so nothing has to be embedded or re-indexed offline and all processing stays local. The trade is scope: one database per project, and the full benefit comes from opening and closing the session properly rather than calling tools ad hoc.
A context server dedicated to one project. It opens a conversation with a snapshot of the codebase, takes updates as files and messages accumulate, answers targeted questions about the project, records milestones, and closes the session with what it learned — all held in a SQL database of your own.
- A session that opens with architectural context, recent changes and prior conversations, at the depth you ask for
- Context kept current as code changes and new messages come in, including across a shift in topic
- Snippets pulled for one specific question, filtered by entity type or file path
- Milestones recorded when a feature lands or an architectural call is made, with an impact assessment
- A closing pass that extracts learnings and suggests next steps
- Retrieval driven by keywords, relationship graphs and structured metadata rather than embeddings, so there is no model to host
A SQL database of your own — the server is built for Turso and takes its URL and an access token as environment settings. One database per project; the schema is created and maintained on first run. Runs under Node.
One command plus a key — npx -y devcontext@latest, then supply credentials