CtrlK
Jules is an autonomous AI coding agent developed by Google under Google Labs, designed to perform software development tasks asynchronously inside secure cloud virtual machines. Originally announced on December 11, 2024, alongside Google DeepMind's Gemini 2.0 model family, Jules operates independently of developer supervision to analyze codebases, write or modify code across multiple files, run tests, and open pull requests on GitHub for human review. Unlike synchronous coding assistants that require a developer to monitor each prompt and response, Jules accepts a task description, plans its approach, executes the work inside an isolated cloud environment, and delivers a completed diff when finished, allowing engineers to redirect their attention elsewhere during execution.
Jules entered public beta on May 19, 2025, at Google I/O, and reached general availability on August 6, 2025. By that point it had accumulated over 2.28 million site visits during its beta period, and developers had collectively contributed more than 140,000 publicly shared code improvements. The agent is powered by Gemini 2.5 Pro at launch, with Gemini 3 Pro integration arriving in November 2025 and Gemini 3.1 Pro rolling out in early 2026.
Google Labs is an internal incubator within Google that develops and tests experimental products before broader release. It has served as the launch vehicle for a range of AI-native tools, including NotebookLM and various Gemini-powered experiments. Jules was conceived as part of a broader Google effort to build "agentic" AI systems capable of completing multi-step tasks without continuous human prompting, a strategic priority articulated explicitly in Google's December 2024 announcements around Gemini 2.0.
Google positioned the December 2024 Gemini 2.0 announcement as the beginning of an "agentic era" in AI, contrasting with earlier models that primarily responded to individual prompts. Jules was presented as a concrete product embodiment of that vision, specifically targeting the software engineering workflow. The name Jules has no publicly documented origin from Google; the agent is referred to simply as Jules throughout official documentation and marketing materials.
The broader competitive landscape that motivated Jules included Devin (AI software engineer) from Cognition AI, which had attracted substantial industry attention earlier in 2024 as an early "AI software engineer," as well as GitHub Copilot's expanding agentic features and the emerging class of autonomous coding tools. By placing Jules under Google Labs and tying it directly to the Gemini model family, Google sought to leverage its model infrastructure advantages in the growing market for autonomous developer tooling.
On December 11, 2024, Google announced Jules as part of a broader Gemini 2.0 event. The announcement described Jules as an "experimental AI-powered code agent" built on Gemini 2.0 that integrates with GitHub workflows. At launch, Jules was described as being able to analyze complex codebases, implement fixes across multiple files, and prepare pull requests autonomously while a developer sleeps or works on other tasks.
Access was initially limited to a select group of trusted testers, with interested developers invited to sign up on a waitlist at labs.google.com/jules. Google characterized the project as exploratory, consistent with its approach to other Google Labs experiments. During this initial phase, Jules was powered by Gemini 2.0 Flash and handled Python and JavaScript codebases.
The December announcement coincided with the unveiling of Gemini 2.0 Flash, which Google described as twice as fast as Gemini 1.5 Pro and designed specifically to serve as the foundation for agentic AI experiences. Jules was cited as one of the first Google products to deploy Gemini 2.0 in a production-like setting.
On May 19, 2025, Jules entered public beta at Google I/O 2025, removing the waitlist requirement and making the agent freely available to any developer worldwide in regions where Gemini models are accessible. This release upgraded the underlying model from Gemini 2.0 Flash to Gemini 2.5 Pro, substantially increasing the agent's reasoning depth and long-context capabilities.
The Google I/O announcement framed Jules as a direct competitive response to OpenAI's Codex (OpenAI), which OpenAI had announced in May 2025 as its own cloud-based autonomous coding agent. Google emphasized Jules' asynchronous execution model and deep GitHub integration as distinguishing characteristics. During the public beta, usage was free with unspecified daily limits, and Google indicated that pricing would be introduced after the beta concluded.
On August 6, 2025, Jules exited public beta and launched with a tiered pricing structure tied to Google's existing Google AI subscription plans. The general availability release introduced several new capabilities including Gemini 2.5 thinking mode integration, higher concurrent task limits, multimodal input support, environment snapshots for reusable setup configurations, and direct integration with GitHub Issues.
By the time of general availability, the beta period had generated 2.28 million site visits, with 45 percent of traffic arriving from mobile devices. Google reported that thousands of developers had submitted tens of thousands of tasks, resulting in over 140,000 code improvements shared publicly. Kathy Korevec, director of product at Google Labs, stated in interviews that developer engagement during beta gave Google significant confidence in Jules' long-term trajectory.
On October 2, 2025, Google announced Jules Tools, a lightweight command-line interface that allows developers to start, stop, and verify tasks directly from their terminal alongside existing shell workflows. Simultaneously, Google opened access to the Jules API in alpha, enabling programmatic task submission, integration with third-party tools such as Slack and Linear, and embedding of Jules into continuous integration pipelines.
The Jules API is organized around three core concepts: sources (input repositories or contexts), sessions (continuous units of work within a specific context), and activities (individual units of work within a session). The Jules GitHub Action, available in the google-labs-code/jules-action repository, allows CI/CD workflows to trigger Jules tasks directly from GitHub Actions pipelines.
On November 19, 2025, Google announced that Jules would be powered by Gemini 3 Pro, beginning with Google AI Ultra subscribers. Google AI Pro users received access in the days that followed. The upgrade brought improved reasoning, stronger intent alignment, enhanced context retention across multi-step workflows, and better performance on complex zero-shot tasks compared to Gemini 2.5 Pro.
By January 30, 2026, Gemini 3 Flash became the default base model for all Jules users, replacing Gemini 2.5 Pro even on the free tier. Gemini 3.1 Pro rolled out to Google AI Pro users in March 2026.
Jules' defining architectural characteristic is its use of isolated Google Cloud virtual machines for code execution. When a developer assigns Jules a task, the agent clones the specified GitHub repository into a fresh VM instance. This approach provides strong isolation guarantees: Jules cannot access files or data outside the cloned repository unless explicitly granted permissions, and the VM is torn down after the task completes. Google describes the execution environment as "private by default" and confirms that Jules does not train on private code submitted to it.
The VM-based architecture enables Jules to install dependencies, run test suites, execute build pipelines, and interact with the file system in the same way a human developer would on a local machine. This is a significant capability relative to language-model-only tools that simulate code execution without actually running it, because it allows Jules to discover failing tests, verify its own patches, and iterate based on real outputs.
Jules uses the Gemini model family for planning, code generation, and reasoning. At general availability, the primary model was Gemini 2.5 Pro, which provides long-context reasoning across up to approximately one million tokens. This context window is particularly valuable for large codebases, allowing Jules to ingest substantial portions of a repository before formulating a plan.
Gemini 2.5 Pro's "thinking" capability, integrated at general availability, allows Jules to engage in extended internal reasoning before generating code or a plan. This feature is distinct from visible chain-of-thought output: the thinking process runs internally and informs the quality of the generated plan and code, but developers see only the resulting plan and proposed changes.
With the Gemini 3 upgrade in November 2025, Jules gained enhanced agentic workflow performance, particularly for multi-stage tasks that require maintaining coherent context across many sequential steps. Gemini 3 Pro's improvements in coding benchmarks translated directly to Jules' task success rates on complex refactors and feature additions.
Jules operates in a fundamentally asynchronous mode. A developer submits a task description, optionally reviews a proposed plan, and then disconnects. Jules runs to completion in the cloud VM, generates a diff, and opens a pull request or provides a review link when finished. Execution time varies widely by task complexity, ranging from minutes for small bug fixes to over an hour for broad refactors.
This execution model contrasts with synchronous tools like Claude Code and Cursor (code editor), which maintain a live interactive session in the developer's terminal or IDE. The synchronous approach allows real-time course correction and is better suited to exploratory or ambiguous tasks, while Jules' asynchronous model is better optimized for well-defined, batch-able work that does not require active developer participation.
Developers can assign tasks to Jules through three primary methods: the Jules web interface at jules.google, GitHub Issues labeled with the "jules" label, or the Jules API. All three methods route to the same underlying execution pipeline. The web interface provides a browser-based diff viewer and chat-like interaction for reviewing Jules' proposed plan before execution begins.
When a task is submitted, Jules performs an initial analysis of the repository to understand the codebase structure, relevant files, and existing patterns. It then generates a multi-step plan articulating the changes it proposes to make, the order of operations, and any dependencies or risks it has identified. The developer can review this plan before execution begins and modify it if needed, maintaining human oversight at the critical planning stage.
Jules introduced interactive planning during its public beta period, allowing developers to engage in a dialogue about the proposed plan before committing to execution. A developer can ask Jules to reconsider an approach, add constraints, or focus on a specific subset of the proposed changes. Once the plan is approved, Jules proceeds to execution without further interaction required.
The Planning Critic agent, added in January 2026, introduced an automatic review pass over generated plans to catch logical errors, missed cases, and overly ambitious scope before execution begins. Google reported a 9.5 percent reduction in task failure rates attributable to the Planning Critic at rollout.
Upon completing a task, Jules opens a pull request on GitHub containing the completed changes. The PR includes a description generated by Jules explaining what was changed and why, making code review easier for human engineers who did not observe the execution. The commit authorship feature introduced in February 2026 allows teams to configure how commits are attributed: as Jules only, as co-authored by Jules and the assigning developer, or attributed to the user.
GitHub Actions integration supports a feedback loop for CI/CD pipelines. The CI Fixer feature, introduced in February 2026, allows Jules to automatically detect and fix GitHub Actions failures on its own pull requests, reducing the manual back-and-forth that previously occurred when Jules-generated code failed automated checks.
Environment Snapshots, introduced at general availability in August 2025, allow developers to save dependency installation scripts and environment configurations that Jules reuses across tasks on the same repository. This reduces the startup overhead of each task by avoiding repeated package installation, which can represent a significant portion of total execution time for projects with complex dependency trees.
Jules is available under three access tiers as of 2025 to 2026.
The free tier provides 15 tasks per rolling 24-hour period and up to 3 concurrent tasks. It includes access to Gemini 2.5 Pro (later Gemini 3 Flash as the base model) and is intended for developers evaluating Jules on real projects without a financial commitment.
Jules in Google AI Pro is included with the Google AI Pro subscription at $19.99 per month. It provides 100 daily tasks and 15 concurrent tasks, with access to Gemini 3 Pro. Google positions this tier as appropriate for daily coding workflows at higher intensity.
Jules in Google AI Ultra is included with the Google AI Ultra subscription at $124.99 per month (as of August 2025; the price was noted at $249.99 in some earlier reports and was subsequently adjusted). It provides 300 daily tasks and 60 concurrent tasks with priority access to the latest models, including Gemini 3 Pro and Gemini 3.1 Pro as they are released. Google targets this tier at power users and teams running multi-agent or high-volume automated workflows.
Google also announced a student access program in which eligible college students can receive a free year of Google AI Pro, granting them Jules Pro access as part of that subscription. Paid plans are accessed through Google One subscriptions, and as of 2025 and early 2026, they are available only to individual Gmail account holders, with enterprise and Google Workspace account support under development.
Task limits reset on a rolling 24-hour basis rather than at a fixed daily cutoff, providing more flexibility for developers who work in bursts at irregular hours.
Jules Tools is a command-line interface released in October 2025 that integrates Jules into terminal-based workflows. It supports starting new tasks, monitoring task progress, stopping tasks, and retrieving completed task results without requiring a browser session. The CLI is designed for developers who prefer to manage Jules tasks alongside shell scripts, Makefiles, or CI/CD pipeline definitions rather than through the web UI.
The Jules GitHub Action, published in the google-labs-code/jules-action repository, allows Jules to be triggered from GitHub Actions workflows. Common configurations include scheduled security scanning pipelines that automatically submit newly opened issues to Jules for triage or remediation, performance improvement agents that run on a schedule against the codebase, and automated dependency update workflows that assign Jules to bump versions and verify test compatibility.
The Jules API, released in alpha in October 2025, exposes Jules' task execution capabilities programmatically. The API supports creating and managing sources (repository connections), sessions (work contexts), and activities (individual tasks). Practical integrations documented by Google include triggering Jules from bug reports filed in Slack, embedding Jules into issue trackers such as Linear, and chaining Jules tasks as steps in automated software delivery pipelines.
The API is described as experimental and alpha, meaning interface specifications and authentication mechanisms may change before stabilization. Google has documented example use cases including a Slack-to-GitHub pipeline in which a user reports a bug in a Slack channel, a bot invokes the Jules API, Jules analyzes the relevant code, proposes a fix, runs the test suite, and reports back to the original Slack thread with a link to the generated pull request.
In February 2026, Jules added Model Context Protocol (MCP) server support for a range of third-party developer platforms including Linear, Stitch, Neon, Tinybird, Context7, and Supabase. MCP integration allows Jules to interact directly with these platforms during task execution, such as querying a Neon database schema while writing migration code or reading Linear ticket descriptions as task context.
Jules occupies a specific niche in the competitive landscape of AI coding tools: asynchronous, cloud-hosted, and tightly coupled to the GitHub pull-request workflow. The following comparison covers the most directly competing products as of mid-2025 to early 2026.
| Feature | Jules (Google) | Codex (OpenAI) | Devin (AI software engineer) | Claude Code |
|---|---|---|---|---|
| Interaction model | Asynchronous, cloud VM | Asynchronous, cloud sandbox | Interactive + asynchronous, cloud IDE | Synchronous, local terminal |
| Underlying model | Gemini 2.5 Pro / Gemini 3 | GPT-4o / o3 | Proprietary | Claude 3.5 / Claude 4 |
| GitHub integration | Native PR creation | Native PR creation | Via integrated IDE | Via GitHub Actions / manual |
| Task limits | 15/100/300 per day by tier | Varies by plan | Usage-based (ACU pricing) | Per-token metered |
| Pricing model | Free / $19.99 / $124.99 per month | Subscription-based | $20 base + per-ACU | Per-token via API |
| Code execution environment | Google Cloud VM | Isolated container | Cloud IDE with internet | Local machine / remote container |
| Developer can go offline during task | Yes | Yes | Partially | No |
| Enterprise availability | Under development (2025-2026) | Available | Available (Goldman Sachs pilot) | Available |
| CLI available | Yes (Jules Tools) | No dedicated CLI | No | Yes |
| Mobile access | Yes (browser) | Limited | No | No |
Codex (OpenAI) launched in May 2025, making it a direct contemporary of Jules' public beta. Both tools are asynchronous, cloud-based coding agents that integrate with GitHub and deliver completed pull requests. The primary differences lie in the underlying model family (GPT-4o and o3 for Codex, Gemini for Jules), the execution environment (isolated containers for Codex versus Google Cloud VMs for Jules), and the ecosystem integration (Jules has tighter integration with Google's broader developer tooling including MCP support and the Jules API).
In practical testing by independent developers, Jules and Codex have shown comparable performance on routine tasks such as bug fixes and test writing, with Claude Code generally outperforming both on complex or ambiguous tasks. Codex's isolation model cuts off internet access after an initial bootstrap phase, while Jules uses Gemini's continuing network access for research during task execution.
Devin (AI software engineer) from Cognition AI is Jules' most prominent predecessor in the autonomous coding agent space. Devin offers a cloud-hosted IDE with terminal access, browser capabilities, and the ability to engage in interactive planning sessions that allow developers to collaborate with the agent in real time. Devin's pricing follows an ACU (Agent Compute Unit) consumption model starting at approximately $20 for a base allocation of 9 ACUs, making cost estimation less predictable than Jules' flat-rate tier model.
Devin has demonstrated stronger autonomous performance on complex engineering tasks and has attracted enterprise adoption, including a reported pilot at Goldman Sachs. Jules, by contrast, is more tightly scoped to the GitHub pull-request workflow and is generally considered more appropriate for well-defined, routine development tasks rather than open-ended engineering projects. Jules' free tier and flat-rate pricing give it an accessibility advantage for individual developers, while Devin's interactive model appeals to teams that want to observe and guide the agent in real time.
Claude Code from Anthropic, released in February 2025, is a synchronous coding agent that runs in the developer's local terminal. It uses the Claude model family and maintains a live interactive session in which the developer can observe and redirect the agent's actions in real time. Claude Code excels at exploratory, ambiguous, or architecturally complex tasks where real-time course correction adds substantial value.
Jules and Claude Code are most directly compared in terms of their interaction model. An independent benchmarking test by developer Jon Atkinson assigned both tools an identical codebase task. Claude Code completed the core task in 12 minutes with one minor follow-up fix; Jules took approximately one hour but successfully installed dependencies and passed the test suite, though it did not complete all requested changes. Atkinson graded Jules a C ("promising potential") and Claude Code a B ("reliable but imperfect"). By Q2 2026, industry practitioners had converged on a pattern of using Claude Code for interactive development and Jules for long-running refactors or test backfill that can be delegated and left to execute overnight.
The two tools are architecturally complementary: Claude Code requires the developer to remain engaged during execution and has full access to the local development environment, while Jules accepts delegated work and returns results hours later with no local environment dependency.
Jules attracted substantial developer interest during its public beta, accumulating 2.28 million site visits with 45 percent mobile traffic, an unexpectedly high share that Google cited as evidence of demand for coding assistance on non-traditional developer platforms. The 140,000 publicly shared code improvements contributed during the beta period provided Google with a visible record of Jules' real-world utility.
Practitioner reviews characterized Jules as an "extremely capable junior developer" suited to routine, well-scoped tasks including bug fixes, test coverage expansion, dependency version bumps, and small feature additions. EngineLabs and similar developer publications noted that Jules consistently saves meaningful time on these categories of work while requiring human oversight to ensure correctness and alignment with project conventions.
The launch of the Jules API and CLI in October 2025 expanded reception beyond direct user testing to include infrastructure and platform engineers interested in embedding Jules in automated pipelines. Hacker News discussions around the API release highlighted interest in self-hosted bug triage pipelines and automated dependency maintenance workflows.
Early reviews noted that Jules, at the time of its May 2025 public beta launch, exhibited characteristics of an early-stage product. Performance on ambiguous tasks was inconsistent, and the daily task cap at the free tier limited the ability to run extended experiments. The practical task limit of 15 tasks per day on the free tier was a common friction point noted in user discussions.
TechRepublic, covering the public beta, observed that Jules "can assist but does not replace developers," a framing that Google itself endorsed. The agent performs best on tasks with clear acceptance criteria and well-defined scope, while it struggles with architectural redesigns, large-scale refactors spanning many interconnected systems, or tasks requiring deep understanding of undocumented internal conventions.
The Hacker News community's response to the Jules API announcement highlighted both the appeal of programmatic access and skepticism about the API's alpha status. Several comments noted that the rolling daily task limits would constrain high-volume pipeline use cases on the free and Pro tiers, and that the absence of enterprise governance features, including SOC 2 attestations specific to Jules, auditable access logs, and admin policy controls, limited adoption in regulated environments.
Jules achieved approximately 51.8 percent on SWE-bench Verified, a benchmark that measures an AI agent's ability to resolve real GitHub issues from popular open-source repositories. This result represents competitive performance relative to peers at the time of the mid-2025 evaluation period, though direct comparisons are complicated by differences in task scope, model versions, and evaluation configurations across tools.
Google has not published official SWE-bench results for Jules. The 51.8 percent figure comes from third-party analyses. Comprehensive, reproducible benchmark comparisons across multiple coding agent tools and diverse repository types remained limited as of early 2026.
Several consistent limitations have been documented across practitioner reviews, user feedback, and independent testing as of 2025 to 2026.
Task scope sensitivity. Jules performs reliably on well-scoped tasks with fewer than approximately 200 lines of code change and clear acceptance criteria. Performance degrades on large architectural refactors, tasks spanning many interconnected systems, or work requiring deep familiarity with undocumented codebase conventions. Multiple reviews describe a pattern of Jules successfully implementing the literal task specification while missing implicit project patterns or introducing subtle regressions in edge cases.
Daily task caps. The free tier's limit of 15 tasks per rolling 24-hour period and the Pro tier's limit of 100 tasks constrain the ability to use Jules for high-throughput automated workflows. Users building pipeline integrations frequently hit these limits, and the ultra tier's 300-task daily cap is required for production-scale automated use.
Concurrency limits. The three concurrent tasks at the free tier, 15 at Pro, and 60 at Ultra cap the number of parallel tasks that can execute simultaneously, limiting team-wide deployment in organizations with many developers simultaneously delegating tasks.
Enterprise governance gaps. As of early 2026, Jules lacks dedicated SOC 2 or ISO attestations at the product level, detailed enumeration of GitHub OAuth scopes in public documentation, exportable audit logs for enterprise compliance requirements, and granular admin policy controls. Google has indicated that enterprise and Workspace account support is under development.
Ambiguity handling. Tasks with ambiguous specifications or references to unfamiliar internal libraries increase iteration cycles. Jules generates a plan based on available context, and if that plan is approved with an incorrect interpretation of the developer's intent, the resulting code may require multiple rounds of correction. Interactive planning partially mitigates this by surfacing the plan for review before execution, but the feedback loop is slower than synchronous tools that allow real-time correction.
Repository access requirement. Jules requires GitHub repository access and does not currently support other version control systems such as GitLab or Bitbucket. Organizations using non-GitHub source control cannot use Jules without migrating or mirroring repositories.