Alp Can Aydin released Tusk, a database client whose AI panel writes the query into a new tab but will not run it
Alp Can Aydin released Tusk on 28 September 2026, a Rust database client for macOS, Windows and Linux where the AI assistant writes a SQL query into a new tab, explains its assumptions, and leaves the "run" step to the user. 148 stars in three days.
Image: alpcanaydin via GitHub
Why it mattersGiving the agent read access to the schema but no write to the database is the small design choice that lets a team of engineers use an AI assistant against production without granting it the ability to drop a table by mistake.
Giving an AI assistant write access to a production database is the request a security review rejects every time, and Tusk is designed to make the request unnecessary. Alp Can Aydin released the Rust database client on GitHub on 28 September 2026, and by 1 October it had 148 stars and 9 forks on the strength of one rule written in the README: "Tusk doesn't let the agent execute SQL."
What the AI panel does, and what it does not
The panel opens with Command L. The agent reads the structure of the tables it needs, writes a query into a new tab, and explains its assumptions in prose. A person then reviews the query and runs it. The README states the limit twice over, in two different places, so it reads as a product decision rather than a short-term restriction the next release will remove. Tusk talks to the agent through the Agent Client Protocol, so the model behind the panel is a user choice between Claude, Codex, Gemini CLI, or a custom agent.
The rest of the application is a keyboard-first desktop client, built in Rust on GPUI. One command palette, a virtualised data grid that keeps smooth scrolling across millions of rows, a SQL editor with completions and diagnostics, structure and index editing with DDL generation, and full undo and redo. The 20 supported databases, as listed in the README: PostgreSQL, Amazon Redshift, CockroachDB, Greenplum, Vertica, MySQL, MariaDB, SQLite, DuckDB, LibSQL, Cloudflare D1, Microsoft SQL Server, Oracle, ClickHouse, Snowflake, BigQuery, Redis, MongoDB, Cassandra, and DynamoDB.
The audience, and the thing the audience was waiting for
A product that gains 148 stars in three days answers a question people were already asking. The common complaint about agent-assisted SQL has been that an assistant that can draft a query will also run one, and the review step the team meant to insist on disappears once someone sees how fast the agent is on its own. Removing the ability to execute from the agent, as a product decision, is what gets a careful reviewer to approve the tool at all.
For a team currently reaching for a frontier model through a chat interface and pasting the output into psql by hand, Tusk is the version of that loop that already has the schema in front of it, writes into a visible tab, and is built on an open protocol so the model is not locked. For a team that was waiting for a reason to let the agent anywhere near the database, the reason is here, and it is "the user still runs the query."
Platform notes
Mac builds require macOS 14 or later on Apple Silicon. Windows builds run on Windows 10 or 11 on x64 and ARM64. Linux builds target x86_64 on Wayland, with Xwayland as a fallback. The licence is MIT and there is no paid tier. The repository was created on 28 September 2026 and has 27 commits as of 1 October.
Source
- alpcanaydin/tusk, GitHub repository, created 28 September 2026
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

