Aperture is an open-source AI code editor that runs five checks on every change before writing it to disk

An AI code editor that proposes a change and leaves the user to check whether it compiles is a worse version of what it is replacing.

Aperture, launched on Product Hunt on 5 October 2026, is a free open-source AI code editor that checks its own work before applying any change. The Composer plans a code modification and stages it as a diff. Before the diff is written to disk, Aperture runs five verification checks: parsing, import resolution, type checking, preview rendering, and test execution. If a test fails after the proposed edit, the agent gets one retry; if that fails, the change stays staged for a human to look at. The project is MIT-licensed, lives at Hankaws/aperture on GitHub, and ships with support for Grok, OpenAI, Anthropic, Gemini, DeepSeek and custom endpoints. The author is Witchayut (@hankaws).

The five checks, in order

Each one answers a different question about the proposed diff:

• Parsing confirms the resulting file is still syntactically valid.
• Import resolution confirms every symbol the diff introduces or renames resolves to something that exists.
• Type checking confirms the diff does not break the project's types.
• Preview rendering confirms a UI change still renders, not just compiles.
• Test execution runs the relevant tests inside a sandboxed browser Worker.

The last one is the piece most agent IDEs do not do today. Tests run in a sandboxed Worker rather than against the user's machine or an external runner, which means they cost nothing to run and do not require the user to configure a test runner before they can use the feature.

The retry policy, and why it matters

The one-retry rule is a design choice, not a limitation. An agent that fails a check and keeps retrying forever is the pattern that burns a budget on a loop that was never going to converge. One retry gives the agent room to try an obvious correction (a missing import, a renamed symbol) and nothing more. If the second attempt fails too, the diff stops being treated as the agent's output and starts being treated as a review item for the user.

That stops two failure modes at once. It stops the agent from reaching a successful parse by writing increasingly nonsensical code, and it stops the user from receiving a successful-looking green pass on code the agent silently rewrote three times.

The two caveats that come with it

The checks only answer questions the project is set up to answer. A project with no tests gets no test gate; a project with no types gets no type-check gate. Aperture is a stronger version of the workflow in a well-set-up repository, and a weaker version in a repository that would benefit from it most.

And the free offer runs on the user's own machine. The demo on the launch page plays recorded runs; a real session on real code needs the user to self-host locally or deploy with their own API keys for the model provider they prefer.

Source

• Aperture on Product Hunt, launched 5 October 2026.