We’ve shipped several updates to Sourcegraph Analytics to give you deeper and more flexible insights into how your teams are using Sourcegraph:

Time filters added to the analytics CSV export API

Sometimes you want to automate large-scale code changes without triggering a merge request, like updating internal tooling, modifying generated code, or experimenting with refactors. Until now, Batch Changes always assumed you wanted to open a merge request.

This update introduces a new type of changeset that doesn’t require a merge request. It simply [...]

When using AI-generated steps (like LLMs) in Batch Changes, results can vary run to run. Until now, re-running those steps meant re-executing the entire batch spec. That’s no longer necessary.

You can now re-execute changeset creation logic for individual workspaces, making it easier to regenerate or revise changesets without restarting from scratch.

Anthropic's new Claude 4 models are now available for Cody users, with the following models available for each plan:

Cody Free and Pro:

We’ve added a new option for site admins to set a default search context for all users on their instance. Previously users could set their own default context used when searching their codebase, but this update now allows admins to set this globally.

This helps teams improve onboarding and search quality by focusing searches on the most relevant parts of the codebase, rather than the entire indexed set of repositories.

File and directory filtering has been added to the compare page in Code Search to make large diffs easier to navigate, manage and share. When comparing branches with thousands of changed files, you can now use the file picker to select specific files or directories to focus on. You can also filter files directly by constructing a URL with multiple file paths, or use a compressed file list for even larger selections.

This update gives you much greater control over which changes you view, helping teams review only the files relevant to them without endless scrolling.

Originally announced as an experimental feature, Cody's agentic context gathering is now generally available across all IDEs and the web.

The context gathering experience evaluates the provided context in a prompt, and fetches any additional context (via Sourcegraph search, terminal, etc.) to provide enhanced, context-aware capabilities. It aims to reduce the need to provide context manually, and at the same time improve the quality of prompt responses.

We've added support for service access tokens for Sourcegraph Analytics, enabling admins to programmatically access user-level analytics data without relying on browser-based authentication. These long-lived tokens are tied to individual accounts.sourcegraph.com users and allow customers to pull analytics data into their own systems securely and continuously.

Previously, customers had to manually extract session credentials to automate analytics access - this update eliminates that workaround and supports a more secure, standards-based integration with customer data pipelines.

Note: This page describes how MCP tools are used within Cody.

If you're looking for the Sourcegraph MCP Server (for connecting external AI agents and applications), see: Sourcegraph MCP server.

Cody now supports MCP tools via its agentic context gathering, enabling automated tool use to gather and inject context directly into Cody's responses.

We've updated the prompt library to make it easier for teams to collaborate using the Organizations feature. Prompts can already be owned by an organization - now, they're easier to find, filter, and manage.

As of Sourcegraph 6.4, customers using self-hosted models through the openaicompatible provider can now configure:

• Custom headers: Add authentication keys or other custom HTTP headers to LLM requests
• Non-legacy completions: Use /chat/completions instead of the older /completions endpoint for Cody's autocomplete queries

We've added a variety of new LLM options from OpenAI and Google for Cody users:

Cody Free:

Smart Apply is a Cody feature that applies high-confidence code edits from chat responses directly into your codebase, streamlining your development workflow. Cody uses a targeted model for Smart Apply that is well-suited for low latency, and the previous model used (Claude 3.5 Sonnet) has been replaced by a new model: Qwen 2.5 Coder.

This update improves the responsiveness of Smart Apply in both VS Code and JetBrains by ~5x while preserving edit quality.

JetBrains users can now try auto-edit, now available as an experimental opt-in in the Cody plugin.

Auto-edit analyzes your recent changes and suggests context-aware edits—like cleaning up code, adding missing pieces, or improving structure. Instead of writing every change yourself, you can apply Cody's suggestions directly in your IDE.

Batch Changes now supports using GitHub fine-grained access tokens as credentials for creating and managing changesets. This support is experimental and is especially useful for customers who prefer not to use GitHub Apps due to their broad repo access, and who also disallow personal access tokens (PATs) because of their wide permissions and lack of expiration controls. Fine-grained access tokens strike a better balance: they’re more secure than PATs and more flexible than GitHub Apps.

They can be used for both user-level and site-admin credentials in Batch Changes. However, note that fine-grained tokens can be more complex to set up and may require admin approval in your GitHub organization.

Site admins can now fully customize the examples shown in the batch spec library - visible when defining a batch spec for a new batch change - using three new GraphQL APIs:

• createBatchSpecLibraryRecord
• updateBatchSpecLibraryRecord
• deleteBatchSpecLibraryRecord

Batch Changes now supports a new bulk action to help teams working in fast-moving codebases: rebase changesets.

This makes it significantly easier for teams to keep changesets up to date. Instead of rebasing each changeset manually, users can now do it all at once with a single action—streamlining workflows and reducing friction when managing large-scale changes.

Developers using Cody chat now have access to two variations of Claude 3.7 Sonnet: the standard model useful for faster, more responsive interactions, and another optimized for reasoning ("extended thinking").

With this update, Claude 3.7 Sonnet can now be used in different contexts depending on whether long-form reasoning is required or for tasks where speed and performance is a priority.

You can now sign commits with SSH keys in Batch Changes - giving enterprise teams a more secure, user-level alternative to GitHub Apps for commit verification. This feature is available in experimental form and currently supports GitHub and GitLab.

To get started, users add a personal access token (PAT) and check a box to enable commit signing. When they do, Sourcegraph generates an SSH key pair, surfaces the public key, and guides users to add it to their code host as a signing key. Site admins can also add PATs, and they function as a global credential to all users who don't set their own.

We’ve improved how Sourcegraph ranks results in the main search bar to deliver more relevant results tailored to each user. When you search for something like data router, Sourcegraph now applies best-in-class ranking strategies that better understand your full query - not just keywords in isolation. In addition, we boost results from repositories you’ve recently contributed to, making it easier to find the code you care about faster.

Flexible code search queries - where you are looking for something, but don’t 100% remember the name - has been improved. Below is a before and after for the query contributors job that highlights this improvement and our use of BM25 ranking.