Sourcegraph unveils Cody 5.1, a free code AI tool that can write entire files and tests

Sourcegraph, a leader in universal code search and AI-assisted software engineering, announced the release of Cody version 5.1 today, a major upgrade to its AI coding assistant. The new version provides Cody with a broader view of code context across repositories and improved automation capabilities, allowing it to generate code, fix bugs and refactor projects with less human intervention.

In an exclusive interview with VentureBeat, Sourcegraph CEO Quinn Slack discussed the new Cody desktop app and its ability to build context for code AI. By allowing developers to point Cody at their local code, the app can better understand the codebase and even write entire tests and files. “Cody now has a deep understanding of codebases that lets developers trust it to write entire files, fix bugs and answer questions about code they’ve never even seen,” he said.

The key enhancements in Cody 5.1, according to Slack, are the ability to understand context across multiple repositories in a codebase and new automation “recipes” that can perform more complex software engineering tasks like optimizing performance, fixing code smells and generating unit tests. Developers get inline access to Cody through a chat interface in their code editors, and Cody can now make changes directly to code.

Cody 5.1 poses challenges for competitors like GitHub’s Copilot, an autocomplete tool that relies primarily on a developer’s current code context. “Copilot was awesome when it was released two years ago, but it hasn’t really changed that much,” said Slack. “Anyone who’s used ChatGPT knows AI could do so much more than a fancy autocomplete.”

Better autocomplete and new recipes

The wider code context comes from Sourcegraph’s strengths as a leading code search and analysis platform, which Cody now taps into. “Cody benefits from 10 years of us building a leading code search engine,” Slack said.

The multi-repository context and more advanced natural language understanding enable Cody to handle ambiguous questions and requests, as well as write idiomatic code by learning patterns across a codebase.

Slack explained that the desktop app generates a local code graph by indexing the code for search and building embeddings for semantic search, enabling the editor to communicate with the app for context when developers use Cody.

“Cody is the first code AI to autocomplete based on context from the entire repository, using embeddings-based semantic search,” Slack told VentureBeat. “This means Cody can generate better code that uses more of your codebase’s own APIs and idiomatic usage patterns, compared to GitHub Copilot and others that only use recent files and open tabs.”

Going beyond autocomplete

Slack also said that Cody 5.1 goes beyond autocomplete and can perform higher-level coding tasks such as writing entire files, tests, docstrings, variable names, release notes, pull request descriptions, optimizing performance, fixing code smells and answering questions about the codebase.

“Cody can explain, write, fix and refactor code using your codebase’s own APIs, documentation, and usage patterns,” said Slack. “This goes way beyond autocomplete or prompt engineering. It’s possible only because Cody supplies context about your own code to a powerful LLM [large language model], so it can perform higher-level coding tasks.”

Cody 5.1 also introduces new features such as inline chat, which allows developers to ask questions and request changes on specific regions of code files; support for JetBrains IDEs, such as IntelliJ, PyCharm, WebStorm; and the Cody desktop app, which makes it easy for individuals to use Cody on their private code in their editor and in a chat UI.

Cody 5.1 is free for developers on both public and private code, with a generous rate limit. Sourcegraph charges only for team/company/enterprise features or for exceeding the rate limit. Sourcegraph Enterprise Server users need to upgrade to version 5.1 to get the new features of Cody.

According to Slack, Cody 5.1 uses more context from the entire codebase and multiple repositories, as well as a more powerful language model, Anthropic Claude, to generate more accurate and consistent code suggestions.

The future of AI in coding

Discussing the role of the open development community in contributing to Cody 5.1, Slack said, “Cody is open source. It’s Apache 2.0, and we’ve received a lot of contributions. I think we’ve got 20 different contributors so far and w’ve got hundreds of people on our Discord.”

He further emphasized the importance of having an open platform and API for developers to get the most out of a product like Cody.

As for the future of AI in coding, Slack envisions a future where AI agents can take multiple steps to improve code without human intervention. However, he believes that building trust between developers and AI is crucial before reaching that stage.

“We’re really excited about [the future of AI in coding],” said Slack. “We’re tracking that really closely. We’re building up to that with Cody as well. Now, we have to proceed cautiously, because at the moment, you have a code AI writing code where no human reviews it, then that’s the point at which the limits to adoption are off.”

Sourcegraph’s vision for AI

The new release is an important step for Sourcegraph in its vision for AI that can automate complex, multi-step software engineering tasks. The company has to proceed cautiously, said Slack, to ensure the AI generates code and outcomes that are appropriate for existing codebases in enterprise settings. But progress toward more advanced automation could significantly boost developer productivity.

“Our approach — more and better context, more powerful LLM — is different from that of other AI code autocomplete tools that optimize for limited context and small models,” he said. “We’re optimistic that this maximal approach will definitively surpass the minimal approach.”

Sourcegraph is a San Francisco-based company that was founded in 2013 by Slack and Beyang Liu. The company has raised $248 million in funding from investors such as Sequoia Capital, Andreessen Horowitz, Insight Partners and Geodesic Capital. Sourcegraph’s annual revenues are estimated to be between $10 million and $50 million, and it has around 160 employees. Sourcegraph’s customers include Amazon, PayPal, Lyft, Uber, Yelp, Cloudflare, Plaid, GE and Atlassian.