Aide is an AI-native , privacy-first IDE built on top of VSCode
We're building Aide, an AI-native IDE designed for engineers to enjoy problem-solving with agents by their side.
We're currently a team of two, based in London, United Kingdom. Between us, we have been tech leads with experience building testing infrastructure at Meta, payment systems for 150k merchants in India, been an ACM ICPC world finalist and Open-Source contributor to critical infrastructure including gRPC and the Bazel build system. We love dev tooling, and enjoy thinking about the future of programming.
Currently building CodeStory. I have worked extensively in the devtooling and AI space for the last 5 years as the tech-lead for testing infrastructure at Meta and as the founding engineer for Findly.ai (YC S22).
Passionate about making developers more productive and want to build for a world where developers have powerful AI agents working together to ship more value.
I love making computers do things. đź‘ľ
Currently building CodeStory, an AI-native IDE designed for engineers to enjoy problem-solving with agents by their side. I've built payments infrastructure to handle 150k merchants and scaled an engineering team from 5 to 35.
Hey all! We’re Sandeep and Naresh from CodeStory. Aide (https://aide.dev) is the IDE for engineers to solve hard problems where changes are not limited to a single file but require iteration across multiple files. We are also open source at https://github.com/codestoryai. The main repositories are the:
When we started building CodeStory last year, we were set on building the best chat and autocomplete experience in the editor, which we believed to be the best use of LLM for code generation.
We quickly (after 6 months) realised that with the increasing trend of LLM coding capabilities, the UX around chat and copilot felt limiting to the real abilities of these LLMs to help the developer accomplish more.
In March this year, while debating what to focus our energy on, we decided to experiment with the idea: “What if LLMs could make edits across multiple files without breaking the logic?” So after a month of hacking, we tested our framework against SWE-Bench Lite and were very surprised (and kind of shocked) that Aide became the SOTA, resolving 43% of the issues.
We instantly knew we had to build the multi-file editing experience into the editor and do it well.
We decided NOT to put energy into building a better “copilot autocomplete” or “chat,” but rather go all in on shipping a better “multi-file editing flow.”
While multi-file edits allow for larger units of work to be done, they have their set of problems:
Lack of control to granularly iterate when there is a small issue within a large change
Inability to roll back to older states when using over a long coding session.
Changes to the local scope, especially when context is not provided exhaustively, break the codebase globally
Context about edits made in a previous editing session is not kept around
To start in this direction, we forked VSCode 15 months ago as the most rock solid foundation to build an editor on top of and put quite a lot of work into solving some of the above stated issues with multi-file editing:
We built a proactive agent which looks at the linter errors (Language server and your own feedback) and proactively comes back with fixes and suggestions for improvements.
Special UX and integration with models for always breaking down changes into small, logical units that are easy to understand and have live previews using decorations in the editor, so changes can be applied easily, in parts or all at once.
Native support for checkpoints and rollbacks: We couldn’t use git, of course, so we hooked into the TextModel implementation in VSCode (which is its in-memory representation of a file) to keep track of file versions across AI changes, while also preserving the undo stack. These checkpoints are exposed in the UI, letting you easily move back and forth when needed without having to worry about uncontrolled changes.
Integrate tightly with the Language Server API on the editor to gather diagnostics from the whole codebase when changes occur, to ensure unintended breakages are quickly caught and propose fixes. For instance, the LLM changing a function signature should also fix the entire call stack.
Maintain a continuous, long session context. Because our editor works entirely locally (no intermediary server interaction, just LLM calls directly from your machine!), the Rust binary shipped with the editor listens to all events and changes made by both the developer and AI to keep a continuous context, so subsequent changes are always better informed.
Though, we, of course, didn’t just stop with multi-file edits.
Even with single-file changes, our Rust binary leverages tree-sitter to structurally understand a file and use a combination of Search-And-Replace style editing with tight editor integration to edit multiple locations within a file in parallel to make faster edits, especially useful in large files. And threw in AST navigation for developers to move faster around them.
To gather context for codebase-wide queries, rather than try to maintain lossy embeddings, the binary builds a repo map to let the LLMs filter down the most useful parts of the codebase per query.
Wherever possible, use the new prompt caching capabilities (also exposed as a first-class citizen in the editor to give developers more control) to cache the most relevant files in a coding session — which largely speeds up response times.
Build a macOS Spotlight style widget to quickly send our edit instructions for the agent to pick up and work while you focus on the code front and centre.
Since everything runs locally on your machine, this is also a more privacy-friendly approach and gives the optional benefit of using your keys & contracts with cloud model providers — and we’re more than happy to speak with companies that are interested in setting this up. If you are someone who loves tinkering with LLMs, all the requests & responses are just stored on a SQLite DB on your machine (and nothing at our end), which is data that could perhaps be used in the longer run to fine-tune or build features to improve AI performance over time within teams.
There is still a ton to do in terms of robustness and capabilities. Despite all the progress, we still feel we’ve only explored about 1% of what coding with AI could look like. For now, we’d love for you to try out Aide at https://aide.dev. To keep things simple, we have a standard subscription plan with a 2-week trial to try out the editor. We’d love to hear your feedback and ideas!
Other Company Launches
CodeStory ✨ is an AI-powered mod of VSCode.
CodeStory is a full-blown AI-powered IDE built to tackle any complex task, so you can accomplish your goals faster.