July 2026 Summaries

GitKraken has introduced Kepler, a free public-preview agentic development environment for Windows, Mac, and Linux that helps teams coordinate AI coding work across multiple repositories and agents. Designed to reduce the management overhead created by parallel AI tasks, Kepler lets developers start work from scratch, existing issues in platforms such as Jira, Linear, GitHub, GitLab, Trello, and Azure DevOps, or pull requests requiring review or feedback resolution. It supports Agent Client Protocol-compatible tools including Claude Code, Codex, Cursor, Gemini, Copilot, OpenCode, and Augment, while offering planning, permission-based, and more autonomous execution modes along with adjustable effort levels for controlling cost and reasoning depth. Tasks can share instructions across repositories and are tracked through list, Kanban, and console views, with remote access available from other browsers or phones. Kepler also integrates with GitKraken Desktop, allowing developers to inspect an agent’s active worktree, review live diffs, and prepare changes for commits and pull requests.

GitKraken Code Review, introduced in the company’s June updates and available through gitkraken.dev, aims to make pull request reviews more focused by combining an AI-generated explanation of a change’s intent with semantic diffing that highlights meaningful modifications. Rather than generating numerous inline AI comments as GitHub-integrated review bots may do, it provides a dedicated review interface where users can comment on lines, mention collaborators, and react without leaving the tool. Because it is web-based rather than tied to GitKraken Desktop releases, features can be updated independently and deployed more quickly. The update also highlights GitBench, a benchmark suite created by developer advocate Chris Griffing that evaluates AI models on Git-specific tasks such as identifying commits to squash, with early tests showing that performance can vary substantially depending on output format and reasoning settings.

GitKraken has launched an official Claude Plugin on the Claude marketplace, allowing users to connect its MCP server to Claude and Anthropic applications without installing a separate CLI tool. The MCP server provides agents with access to GitKraken integrations and features while helping them understand codebases in ways intended to reduce token usage and improve results. GitKraken plans to expand the plugin with additional skills and resources, welcomes community feedback through its GitHub repository, and is developing similar plugins for tools such as Cursor and OpenAI.

GitLens 18 expands the extension from a Git history viewer into a workspace for monitoring parallel development and AI coding agents across VS Code-based editors, including Cursor, Windsurf, and Antigravity. Its updated commit graph displays uncommitted changes from open worktrees alongside agent sessions, while integrated review, compose, and compare panels provide AI-assisted change summaries, commit creation, and branch-level comparisons. New agent-session and tree-map views show which agents are active, idle, or awaiting input and where they are modifying a repository in real time. Visual History now covers files, folders, and entire repositories, helping developers examine code evolution and contributor activity, while branch pinning, branch focus, and a detachable graph window improve navigation and monitoring. GitLens remains free through the VS Code Marketplace, has more than 40 million installs, and can be updated through the editor’s Extensions panel, with a pre-release channel available.

GitKraken Desktop 12 adds agent sessions to its established Git visualization and repository management workflow, allowing developers to launch or work alongside coding agents such as Claude Code, Codex, Gemini, OpenCode, Copilot, and Augment within the same application. Designed for focused work in a single repository, it combines live diffs, commit history, branch and remote views, and the commit graph to help users inspect agent-generated changes before staging, committing, pushing, and opening pull requests. The application can also detect agents editing a worktree outside the client and provide a shortcut directly to the relevant live diff, emphasizing review and developer control. This single-repository approach differs from Kepler, which is intended to coordinate agents across multiple repositories for broader tasks.

Google’s Gemini 3.6 Flash, 3.5 Flash Lite, and 3.5 Flash Cyber models have drawn criticism for trailing leading OpenAI and Anthropic models on broad intelligence and coding benchmarks, despite expectations that Google’s data resources and Transformer research heritage would give it an advantage. Benchmark comparisons indicate that Gemini’s Flash models generally offer lower latency and higher throughput, particularly relative to similarly positioned competitors, but their advantages can be reduced by high token consumption, which raises per-task costs and can negate speed gains for Gemini 3.6 Flash. Flash Lite performs more favorably across combined speed, cost, and intelligence comparisons, approaching an efficient balance for tasks that do not require top-tier reasoning ability. The models may be particularly appropriate for high-volume applications such as document summarization, programmatic workflows, and search-result summaries, where responsiveness matters alongside capability. GitKraken’s GitBench microbenchmark also showed Google models performing strongly on specialized Git-related tasks, illustrating that broad rankings may not capture performance in particular domains. Overall, model selection should weigh intelligence, speed, token use, and cost according to the intended task rather than relying solely on general-purpose benchmark rankings or social-media perceptions.

GitKraken’s Kepler, currently in public preview as version 0.8, is designed to help developers coordinate multiple AI coding agents, repositories, issues, and pull requests without constantly switching among terminals, Jira, GitHub, and agent interfaces. Created from senior engineer Gyo’s frustration with manually launching and tracking agent tasks, Kepler provides a centralized workspace where developers can initiate work on issues, review pull requests, monitor in-progress changes, and remain involved before code is published or merged. Rather than replacing integrated development environments, Kepler is intended to complement them by supporting agent orchestration and human oversight, while IDEs remain useful for reading, validating, and refining generated code.

Merge conflicts create hidden productivity costs by interrupting developers’ focus and requiring them to reconstruct the intent behind competing changes, a burden that is often absent from sprint metrics. GitLens 18.2 introduces AI-powered conflict resolution for merges, rebases, and cherry-picks within VS Code, using its Commit Graph, sidebar, and Command Palette to process all conflicted files in parallel. The tool proposes resolutions, classifies outcomes such as automatic merges, current or incoming versions, deletions, and files needing review, then displays each result as a diff with explanations before any changes are applied. Developers retain final control, can refine individual or complete resolutions through prompts, and can leave uncertain files unresolved. GitLens 18.2 is available through the VS Code Marketplace, with AI-powered resolution for private repositories offered through GitLens Pro.

GitBench is a GitKraken initiative that evaluates how effectively large language models perform Git-specific tasks, using 204 test cases across 17 benchmark categories and results from 43 or more models at different reasoning levels. Created to address the lack of comparable evaluations for features such as commit-message generation, pull-request generation, merge-conflict resolution, rebasing, and commit composition, the project aims to complement broad reasoning and coding benchmarks while reducing incentives for models to optimize narrowly for a small set of tests. Its Python CLI uses Click to run benchmarks and generate JSON, while an Astro-based web application with Recharts and shadcn/ui presents searchable, statically rendered results; OpenRouter provides access to most tested models. Development uncovered several evaluation challenges, including uneven structured-output support, invalid or poorly targeted JSON responses that required more specific schema keys, and weaknesses in string-similarity scoring that were improved through aggregated judgments from multiple lower-cost LLMs. The team also encountered unexpectedly inappropriate and excessively verbose output from Nemotron, requiring content redaction and changes to data-validation workflows. GitBench’s public site presents comparative model performance, cost, and token-use findings, with the creators encouraging others to develop task-specific model evaluations.