The logging half is meant to be invisible — the assistant classifies as it goes and records without interrupting the conversation. The querying half is where it pays off, because the store is an ordinary SQLite database with both a formatted view and raw SQL over the same rows. Every entry carries the reasoning for its classification, so the log answers not only what was flagged but why.
An agentic logging tool. It classifies each interaction by how ordinary it is, records the unusual ones with the reasoning behind the classification, and stores them in a local SQLite database that stays queryable afterwards.
- Classification and logging of unusual turns, with common interactions passed over, unusual patterns recorded, and critical events logged as a priority — `log-query`
- A readable log view with filtering, which is the tool for "what happened this morning" — `read-logs`
- Direct SQL against the store: SELECT on one side, INSERT, UPDATE and DELETE on the other — `read_query`, `write_query`
- Schema-level access to the same database — `create_table`, `list_tables`, `describe_table`
- A defined schema for each entry: session and user ids, interaction type, probability class, the message and response, a context summary and the reasoning
No account and no key. Python 3.10 or higher, MCP 1.2.0 or higher, SQLite3, and `uv`. Clone and run `uv pip install -e .`, then register `uv` in your client config with `--directory` pointing at the clone and `mcp-variance-log` as the command to run. The database is created at `data/varlog.db` relative to the installation, and that directory must exist. The README also supplies the client-preferences text that tells the assistant to log silently as it goes.
Build from source — clone the repository and build it, then point your client at the binary