What Is Jules? Google's Async Coding Agent That Runs While You Sleep

Jules is an asynchronous AI coding agent from Google Labs that clones your GitHub repository into a secure Google Cloud VM, writes a plan, executes multi-file changes, and opens a pull request for review. It is powered by Gemini 2.5 Pro on the free tier and Gemini 3 Pro on paid tiers, launched in public beta on May 20, 2025, and exited beta on August 6, 2025 according to Google's announcement covered by TechCrunch. The free plan allows 15 tasks per rolling 24 hours, positioning Jules as a background-execution alternative to interactive tools like Cursor or Copilot.

Key Takeaways

• Jules is a proprietary, cloud-hosted coding agent — not an editor plugin, not open source.
• It runs asynchronously: assign a task, close the laptop, review a pull request later.
• Free tier: 15 tasks/day, 3 concurrent, Gemini 2.5 Pro. Ultra tier: 300 tasks/day, 60 concurrent, priority access to Gemini 3 Pro.
• Core primitives are Sources (GitHub repos), Sessions (units of work), and Activities (steps within a session).
• A V2 rewrite is in active development with an early-access waitlist; a new logo and octopus-themed rebrand have surfaced in mid-July 2026 previews.
• Access is via jules.google.com, a CLI called Jules Tools, or the v1alpha REST API at jules.googleapis.com.

What Is Jules?

Jules is a proactive, autonomous coding agent built by Google Labs. It was first announced in December 2024, opened to public beta at Google I/O on May 20, 2025, and reached general availability on August 6, 2025. Its purpose is narrow and specific: take a coding task described in natural language, clone the target repository into an isolated Google Cloud VM, produce a step-by-step plan, execute the plan, run tests, and open a pull request against the source branch.

Jules is not a chat assistant and not an autocomplete tool. It does not run inside VS Code or JetBrains. The interaction model is asynchronous — the closest analogues are OpenAI's Codex cloud agent and Cognition's Devin, not Cursor or Copilot. That distinction shapes every other design choice below.

The V2 rewrite currently in development is described by the Jules team, in community posts from testers and Google Labs staff, as "a complete rewrite" that shifts Jules from "a coding agent" toward an "end-to-end agentic product development platform" — one that reads full product context and prioritizes what to build. That framing is still pre-release and unconfirmed in official documentation.

Jules at a Glance

Field	Value
Developer	Google Labs
Type	Asynchronous AI coding agent
Modality	Code (multi-file edits, PRs) with multimodal input support
Model backend	Gemini 2.5 Pro (free); Gemini 3 Pro (Pro/Ultra)
Context window	Not yet published for the agent (relies on Gemini backends)
Free tier	15 tasks/day, 3 concurrent
Paid tier — Pro	100 tasks/day, 15 concurrent — $19.99/month via Google AI Pro
Paid tier — Ultra	300 tasks/day, 60 concurrent — $124.99/month via Google AI Ultra
Availability	General availability since Aug 6, 2025
Interfaces	Web app, Jules Tools CLI, REST API (v1alpha)
Repository support	GitHub (via Jules GitHub app)
License	Proprietary; closed weights

Prices and limits reflect Google's official Jules documentation as of July 2026.

How Jules Works

Jules is built around three primitives documented in the Jules API reference:

• Source — an input repository, typically a GitHub repo, connected through the Jules GitHub app.
• Session — a continuous unit of work initiated by a prompt against a source, similar to a chat session.
• Activity — a single event inside a session: a plan generation, a message, a progress update, a completed artifact.

A typical run has four phases. First, the developer assigns a task through the web app, the CLI, or a labeled GitHub issue. Second, Jules clones the entire repository into a fresh Google Cloud VM. If an Environment Snapshot (Google's term for a saved dependency/install/env-var bundle) exists, Jules reuses it for faster setup. Third, Jules produces a plan, which the developer can review, edit, or reject before execution. Fourth, Jules executes the plan, runs tests, and — if automationMode is set to AUTO_CREATE_PR — opens a pull request.

Two design details matter for AI engineers. Multiple internal agents appear to be involved: security researcher Johann Rehberger, publishing on his blog Embrace The Red in August 2025, described a main planner that hands subtasks to worker agents with distinct tool sets. And Jules has non-trivial tool access, including a view_text_website tool for reading external pages — which the same writeup showed can be abused as a data exfiltration channel during prompt injection from GitHub issues. Anyone connecting Jules to sensitive repositories should treat issue bodies and fetched web content as untrusted input.

What You Can Do With Jules

The concrete workloads Jules is designed for, based on Google's documentation and community reports:

• Backlog burndown: assign small bug fixes or dependency bumps via GitHub issue labels and let Jules ship PRs in the background.
• Test writing: point Jules at an under-tested module and ask for coverage; the agent writes tests, runs them in the VM, and iterates.
• CI Fixer: a named feature that automatically diagnoses and proposes fixes for failing CI runs.
• Multi-file refactors: because the whole repo is cloned, Jules can make coordinated changes across files without the working-set limits of an IDE agent.
• Overnight tasks: the async model is explicitly marketed as work that runs "while you sleep," according to the jules.google.com landing page.

Community sentiment in mid-2026 is mixed. Long-term users on Reddit report the free tier is genuinely useful for "minor bugs" and preparing a foundation before a work session, while others describe latency (up to 40 minutes per small task) and stalled GitHub branch publishes during the beta period. Google says stability improved substantially before GA, but reliability against Claude Code or Codex remains a live debate.

How Jules Compares

Jules occupies a specific quadrant: async, PR-based, no editor integration. That is the same quadrant as Codex's cloud mode and Devin, and a different one from Copilot, Cursor, or Claude Code's interactive modes.

Feature	Jules	Claude Code	OpenAI Codex	Devin
Interaction mode	Async (PR-based)	Interactive + async	Async (sandbox)	Async (cloud)
Editor integration	None (GitHub only)	VS Code, terminal	VS Code, terminal	Browser IDE
Free tier	15 tasks/day	Limited free	Limited free	No
Entry paid price	$19.99/month (AI Pro)	$20/month (Pro)	$20/month (Plus)	$500/month
Model backend	Gemini 2.5 Pro / Gemini 3 Pro	Claude Opus 4.6	GPT-5.3 Codex	Proprietary

Comparative pricing above is drawn from vendor documentation and a third-party comparison writeup; the Codex reorganization context is covered in our earlier analysis of OpenAI's July 9 Codex-ChatGPT merge.

Availability: How to Access Jules

Three access paths exist today:

1. Web app at jules.google.com. Sign in with a Google account, connect the Jules GitHub app to grant repository access, and start assigning tasks. Free tier requires no payment; paid tiers are enabled by subscribing to Google AI Pro or Google AI Ultra. Currently paid plans are only available for individual Google Accounts ending in @gmail.com; Workspace and enterprise upgrade paths are "coming" per Google's docs.
2. Jules Tools CLI, documented on jules.google, for scripting task creation from the terminal.
3. REST API at https://jules.googleapis.com/v1alpha, currently in alpha. API keys are generated in the Jules web app under Settings (maximum three at a time) and passed in the X-Goog-Api-Key header. The API exposes sessions, sessions.activities, and sources resources with methods like create, sendMessage, and approvePlan per the Jules API reference documentation. Because the API is alpha, Google warns that specifications, keys, and definitions may change.

If you want to prototype similar agentic workflows on a different backbone while you wait for Jules API access, general-purpose coding-capable models like OpenAI Codex on kie.ai cover overlapping ground with a stable public API today.

Age requirement: Jules requires users to be 18 or older, which is stricter than some Google One plans — a shared family plan does not grant access to under-18 members.

What We Don't Know Yet

Several open questions matter for anyone planning around Jules:

• V2 release window. The @julesagent team has sent branded merchandise (a recurring octopus/squid motif) to influencers, and posts from TestingCatalog on July 14, 2026 and AshutoshShrivastava on July 17, 2026 describe an imminent V2 with a new logo. An external tester writing on X mid-July 2026 said V2 is "at least a month more" from release. Google has not published an official date.
• V2 feature set and pricing. Community reports describe a rewrite for UI speed, native multi-repo support, and a shift toward end-to-end product development. Model backend, revised limits, and any pricing changes are unconfirmed.
• Enterprise upgrade path. Google Workspace and business accounts cannot yet subscribe to paid Jules tiers; the timeline for enterprise plans is not published.
• Benchmark disclosure. Google has not published Jules-specific SWE-bench Verified numbers for the current agent stack. A third-party writeup cited 51.8% on SWE-bench Verified for a Gemini Flash configuration, but this is not an official Jules benchmark and should not be treated as one.
• Public repo data usage. Google's post-GA privacy update states public-repo data may be used for training while private-repo data is not — teams with sensitive OSS work should read the current privacy notice directly.

Frequently Asked Questions

What is Jules used for?

Jules is used to hand off coding tasks — bug fixes, feature additions, test writing, dependency bumps — that run asynchronously in a Google Cloud VM and return as a GitHub pull request. It is not an inline autocomplete tool; developers assign work and review a diff later. The typical workload is "backlog items I don't want to context-switch into," not real-time pair programming.

Is Jules free to use?

Jules has a free tier that allows 15 tasks per rolling 24 hours with 3 concurrent tasks, powered by Gemini 2.5 Pro. Paid access is bundled into Google AI Pro at $19.99/month (100 tasks/day, 15 concurrent) and Google AI Ultra at $124.99/month (300 tasks/day, 60 concurrent). Paid tiers get higher or priority access to Gemini 3 Pro.

What model powers Jules?

Jules runs on Gemini 2.5 Pro on the free tier. Paid tiers get higher or priority access to Google's latest model, starting with Gemini 3 Pro according to the official documentation. The agent orchestrates multiple internal roles — planner and worker agents — on top of that backbone, with distinct tool sets per role.

How is Jules different from GitHub Copilot or Cursor?

Jules is asynchronous and repo-scoped rather than editor-embedded. Copilot and Cursor work synchronously inside an IDE, watching each keystroke; Jules runs in a background Cloud VM, so you can close your laptop and return to a finished pull request. The tradeoff is latency: individual Jules tasks take minutes, not milliseconds, but they can span multiple files without a human in the loop.

Is Jules open source?

No. Jules is a proprietary Google Labs product. Its weights, agent orchestration, and Cloud VM harness are not publicly available; only a REST API and CLI are exposed for programmatic use. This is a significant contrast with open coding stacks and matters for teams with air-gapped or on-prem requirements.

How do I access the Jules API?

The Jules REST API is in v1alpha at https://jules.googleapis.com/v1alpha. Developers generate an API key in the Jules web app Settings page (maximum three keys) and pass it in the X-Goog-Api-Key header. The API exposes Sources, Sessions, and Activities resources with methods for creating sessions, sending messages, approving plans, and listing activities.

When was Jules released?

Jules was first announced as a Google Labs experiment in December 2024, entered public beta on May 20, 2025, and exited beta on August 6, 2025. A V2 rewrite is currently in development with an early-access waitlist at jules.google.com/earlyaccess, targeted at teams with live products.

What to Watch Next

Three concrete signals will drive the next update to this page. First, the V2 launch itself — a new logo, an early-access-to-general transition, and any pricing or limit changes. Second, an official model card or SWE-bench Verified number for the current Jules agent stack, which Google has not yet published. Third, the arrival of enterprise and Google Workspace upgrade paths, which are the current gating factor for team adoption.

Building similar async coding or agentic workflows? On kie.ai you can try OpenAI Codex, Claude Sonnet 5, and Kimi K3.