Guide
Google Jules: a guide to plans, tasks, and scheduled runs
By Atlas AI Resources · Published November 13, 2026
Jules is Google's asynchronous coding agent. You give it a GitHub repository and a task, it writes a plan, and then it works in a cloud virtual machine while you do something else. This page starts with our own advice for a first setup, then summarizes what the Jules docs say about tasks, environments, scheduled and suggested tasks, the command line and API, and plan limits.
1. Our advice: a first safe setup
- Connect only the repositories you mean to use. The GitHub flow lets you choose all repositories or specific ones. Pick specific ones; you can widen access later from GitHub's application settings.
- Read the plan; don't let the timer approve it. In the web app a plan you walk away from gets approved automatically after a while, and API sessions approve plans by default. For anything beyond a small fix, stay for the plan or, if you use the API, set requirePlanApproval to true.
- Write a setup script and snapshot it. A tested install and test command saves every later task from guessing how your project builds.
- Keep secrets out of the repository. The VM has internet access and Jules reads non-code files too, so a token in a config file is a token the agent can see.
- Scope tasks like tickets. "Add a test for this function" works; "make this better" doesn't. If a task is too big to review in one sitting, split it.
- Start scheduled tasks small. One recurring job on one repository, checked for a couple of weeks, before you add more. Judge it by the changes you actually merge.
2. Per the docs: how a task runs
- Sign in and connect GitHub: sign in at jules.google.com with a Google account, accept the privacy notice once, and connect GitHub, choosing all or specific repositories.
- Start a task: pick a repository and branch (the default branch is preselected), write a specific prompt, optionally add setup scripts, and select Give me a plan. You can attach PNG or JPEG images such as mockups or screenshots at task creation, up to 5 MB in total; images can't be added to follow-up prompts yet and aren't committed to your repository.
- Review the plan: after cloning the repository and installing dependencies, Jules shows its plan as expandable steps with any assumptions. You can revise it in chat and then select approve plan. If you navigate away, the plan is auto-approved on a timer.
- Watch and steer: an activity feed logs each step, outputs, errors, and questions, with a mini diff per file and a full diff editor. You can send feedback mid-task or pause Jules at any time.
- Finish: the summary lists files changed, runtime, lines changed, and a branch name and commit message. Publish branch or Publish PR pushes the work to GitHub. Jules is the commit author; if Jules publishes the PR it appears as the PR's creator, and if you open the PR yourself from the branch you're the author.
- From an issue: adding the label "jules" (any case) to a GitHub issue starts a task, as long as the Jules GitHub app can access the repository. Jules comments on the issue and links the pull request when it's done.
- AGENTS.md: Jules looks for an AGENTS.md file in the repository root and uses it to understand the code and plan better.
- Failures: Jules retries a failed task automatically, then marks it failed and notifies you. The docs name broken setup scripts and vague prompts as common causes.
3. Per the docs: environments and snapshots
- Each task runs in a short-lived Ubuntu VM with Node.js, Bun, Python, Go, Java, Rust, Docker, and common build tools preinstalled.
- For simple projects Jules works out the setup itself, using AGENTS.md or the README as hints. For complex ones you add a setup script under the repository's Configuration, in the Initial Setup window.
- Run and Snapshot runs the script and, if it succeeds, saves a snapshot of the environment that later tasks on that repository start from.
- Long-running processes like dev servers or watch scripts aren't supported in setup scripts; use discrete install and test commands.
- The FAQ says the VM has internet access, that you're responsible for the code you run, and that Jules doesn't train on private repository content.
4. Per the docs: scheduled and suggested tasks
- Scheduled tasks turn any task into a recurring one. Choose Scheduled Task from the Planning dropdown in the task input, set a frequency such as daily or weekly, and submit. The Scheduled tab shows the last and next run. Editing isn't supported yet, so changing one means deleting it and creating a new one.
- The docs offer three scheduled-task templates (performance, design, and security) that the team behind Google's Stitch design tool uses.
- Suggested tasks are an experimental feature. Turn on Proactivity for a repository, up to five repositories, and Jules scans it periodically and proposes fixes. For now it looks for #TODO comments that describe resolvable work. You review each suggestion and select Start, or dismiss it, and new suggestions also arrive by email.
5. Per the docs: CLI and API
- Jules Tools is a command-line tool installed with
npm install -g @google/jules. Afterjules login,jules remote newstarts a cloud session (it can infer the repository from your current folder, and--parallelstarts several sessions on the same task),jules remote listshows repositories or sessions, andjules remote pullbrings back a finished session's changes. Runningjulesalone opens a terminal dashboard with a side-by-side diff viewer. - The REST API is an alpha release, so its specifications and keys may change. You create up to three API keys in the web app's Settings and pass one in the
X-Goog-Api-Keyheader. Its resources are sources (connected repositories), sessions (one task with a prompt and a source), and activities (steps inside a session). - API sessions approve their plans automatically unless you set
requirePlanApprovalto true, and they only open a pull request on their own whenautomationModeis set toAUTO_CREATE_PR. Publicly exposed keys are disabled automatically.
6. Per the docs: plans and limits
The plans page lists three tiers. Limits count over a rolling 24 hours, and the docs say they may change.
- Jules (no cost): 15 tasks a day, 3 at once.
- Jules in Pro (included with Google AI Pro): 100 tasks a day, 15 at once, with higher access to the latest model.
- Jules in Ultra (included with Google AI Ultra): 300 tasks a day, 60 at once, with priority access to the latest model.
Paid plans currently go through a Google AI subscription on personal Google accounts ending in @gmail.com, and there's an interest form for business users. Family plan members each get their own limits, and Jules requires users to be 18 or older. At the limit, the new task button is disabled, but you can still review and manage existing tasks.
7. How this compares
Jules follows the same basic pattern as other cloud coding agents: an isolated machine, a prepared environment, and a branch or pull request that waits for a person. Its plan-first step is the part that stands out. See our OpenAI Codex Cloud guide, GitHub Copilot cloud agent guide, and Cursor Cloud Agents guide. Whichever you use, keep the real gate in your repository settings (branch protection and required reviews), not in the prompt.
Related: OpenAI Codex Cloud guide · GitHub Copilot cloud agent guide · Cursor Cloud Agents guide · Claude Code guide · Scheduling AI agent routines · Keeping AI agents safe from prompt injection · Parallel agent campaigns · Devin guide