DBQuill

Asking your database in plain language

DBQuill is an open source, local first Windows AI database agent for querying, understanding, visualizing, and safely updating SQLite, MySQL, and PostgreSQL. Its core behavior is to turn a request into a typed, reviewable database operation before anything reaches the database. Read paths are physically read only and bounded, while write paths stop at validation and a change preview until the user explicitly confirms them. The README frames the experience around four ideas. You ask naturally: inspect schemas, search rows, calculate grouped metrics, compare periods, and continue with follow up questions. You see useful results: switch between tables and colorful charts without losing the underlying evidence. You write deliberately: choose a target table, fill a form based on its real columns and sample shape, preview the change, then confirm. You keep control locally: database credentials, model profiles, sessions, audit records, and uploads stay outside the source tree. The quick start on Windows needs Windows 10 or 11 on x64, CPython 3.12, the Microsoft Edge WebView2 runtime, and Git. A clone followed by a single script creates an isolated virtual environment, installs a hash locked dependency set, runs diagnostics, and opens the desktop app. Any OpenAI compatible text model endpoint can be added in settings, so no specific vendor is required.

Verified data sources

The README is explicit about what has been verified rather than暗示 broader support. For SQLite, the path is complete: schema, retrieval, metrics, charts, semantics, scheduling, and confirmed writes make up the minimum viable product. CSV and .xlsx files import into a local SQLite database, with the caveat that legacy .xls is not supported. MySQL 8.4 supports schema, primary and foreign key discovery, bounded queries, grouped metrics, timeouts, and physical read only enforcement; controlled insert, update, and delete are opt in, and a live vendor write matrix is still pending. PostgreSQL 17 covers the same scope as MySQL 8.4, with the same opt in controlled writes and pending matrix. The source setup is continuously checked on a clean, GitHub hosted Windows environment, and the README stresses that a release badge or source archive is not a claim of a signed native installer; the current distribution is source first and requires Python 3.12. This honesty about boundaries is a notable part of the project's tone. A user evaluating DBQuill can see at a glance which databases are safe to write to today and which are read only pending further work, rather than assuming every advertised database behaves identically. Stating the pending write matrix for MySQL and PostgreSQL is honest about where the tool is today. A team that needs safe writes on those databases knows to enable the controlled path and test it, rather than assume the SQLite experience carries over unchanged.

Security and architecture

The architecture treats the model as a planner input, not the authority. Deterministic schema operations and proven relational plans bypass model generated SQL whenever possible, and every candidate still crosses the same gates: authorization, single statement, row limit, semantic, and execution. The desktop client talks only to a loopback aiohttp service protected by a local token and same origin checks. SQLite reads use physical read only connections with query_only, and remote reads use a separate read only session even when controlled writes are enabled. Security and privacy rules are strict. Queries are read only by default, single statement, and row bounded. Writes require validation, a change preview, and explicit confirmation. Remote writes require an explicit controlled write connection, only data manipulation language is enabled, and remote data definition language stays blocked. Scheduled natural language work cannot approve writes automatically. Audit records keep controlled metadata and hashes, never raw prompts, SQL, credentials, or result rows. Model credentials live in an ignored local config and are never required in repository files. The project gate in CI checks repository hygiene and credentials, compiles critical modules, runs the full security and functional regression suite, validates a fixed offline evaluation set, checks frontend JavaScript, and verifies the recorded source fingerprint. DBQuill is released under the MIT license.

Editorial conclusion

DBQuill is released under the MIT license and targets Windows 10 and 11 with CPython 3.12 as the runtime.