Google Jules review (2026): the free tier is the product
A Google Jules review for engineers: the label-to-PR loop, the 15-task free tier, what Gemini-only and no Linear trigger mean, and who should pick it.
Jules is Google's asynchronous coding agent, and the shortest fair review of it is that it gives away more than anyone else in the category. Fifteen tasks a day, free, on a real agent that clones your repository into a cloud VM, runs your tests, and opens a pull request. Everything else about Jules follows from Google's decision to bundle it into consumer AI plans rather than sell it to teams: the models are Gemini only, the trigger is GitHub only, and the paid tiers are sold the way Gemini is sold. It is also better than its reputation in a couple of places that matter.
Checked in early September 2026 against Jules' own docs, changelog, and Google's subscription pages, linked throughout.
Sinatra is our product and it competes with Jules on the label-a-GitHub-issue loop, so this is a competitor writing a review. We have tried to write the things Jules wins as plainly as the ones it loses, and the free tier is one it wins.
What Jules is
Jules runs on Google's infrastructure rather than yours. Per its FAQ, "each task runs in a fresh virtual machine where Jules clones your repo, installs dependencies, and makes changes based on your prompt." The VM has internet access and, per the changelog, 20 GB of disk. You describe the task in the web app, or you label a GitHub issue, and Jules writes a plan, executes it, and hands back a diff you can publish as a branch or a pull request.
It reads an AGENTS.md file in the root of your repository for standing instructions, and it has a memory feature, switched on per repository, that saves your corrections and preferences across tasks.
How a task runs
The GitHub path is the one to care about. Jules' changelog put it plainly when the feature shipped: "Add the label 'jules' to any GitHub issue to start a task in Jules. That's it, label on, task live." The running tasks docs add that the label is case insensitive and that Jules posts a link to the pull request on the issue when it is done.
Between label and pull request, the loop is plan first, then code. Jules drafts a plan and you can approve it before anything changes, which is the most cautious default in this category. Then it works in the VM, where it can install dependencies and run tests. For projects with a non-trivial setup, the environment docs let you supply a setup script that runs before the task, with the caveat that long-running processes like dev servers are not supported there.
Review is where Jules is better than most people expect. Since September 2025, per the changelog, Jules reads comments on its pull requests, adds a 👀 reaction to each one it has seen, and "based on your feedback, Jules will then push a commit with the requested changes." That is the default. A Reactive Mode setting narrows it to comments that mention @Jules. Most agents in this market need a mention every time; Jules acts on the review itself.
What it costs
Jules meters tasks, not tokens, and there is no model bill on top. From the usage limits page:
| Plan | Tasks per day | Concurrent |
|---|---|---|
| Free | 15 | 3 |
| Jules in Google AI Pro | 100 | 15 |
| Jules in Google AI Ultra | 300 | 60 |
The paid tiers come with Google's consumer AI subscriptions rather than a Jules plan of their own. On Google's US subscriptions page, Google AI Pro is $19.99 a month and Google AI Ultra starts at $99.99, with a higher Ultra tier at $199.99. Jules is one line item among many on those plans, so whether the upgrade is worth it depends on whether you wanted the rest of the bundle.
For an individual, this is the cleanest pricing in the category. There is nothing to forecast and no provider account or key to manage, and fifteen tasks a day covers most people's real delegation volume without spending a cent.
Where Jules is strong
The free tier, first. Three times Sinatra's daily allowance and the same agent paying users get. If you are evaluating whether background agents work on your codebase at all, Jules is the cheapest possible experiment.
The review loop. Acting on pull request comments by default, with a visible acknowledgment per comment, is the right design and Jules had it before most of its competitors.
The surface area, which has grown well past one loop. Scheduled tasks run recurring maintenance without re-prompting. Suggested tasks scan a codebase and propose work. The Render integration watches for failed deploys, reads the logs, and pushes fixes to Jules' own pull requests. There is a CLI, Jules Tools, and a REST API for wiring it into anything else. And MCP support, added in February 2026, connects Linear, Neon, Supabase, and a few others as tools the agent can call mid-task.
The models. Jules runs the Gemini 3 family, with Gemini 3.1 Pro gated to Google AI Pro users per the changelog, on Google's own infrastructure, which is the shortest path to a long-context model for a large repository.
Where it stops
Gemini only, and no key. There is no bring-your-own-key path in the docs, and no way to run a Jules task on Claude or GPT. If your team has a view about which model handles its hard tickets, Jules does not let you act on it.
GitHub only. The one documented tracker trigger is the jules label on a GitHub issue. The MCP connection to Linear lets the agent read and update Linear during a task, which is useful, and it is not a product manager assigning a ticket. A team whose backlog lives in Linear has no way to hand work to Jules from there.
Consumer plans. Jules' own usage-limits page says paid plans come through a Google AI Plans subscription that is "currently available only for individual Google Accounts (ending in @gmail.com)," and Google's subscription page asks for "a personal Google Account that you manage on your own." A company that wants to roll Jules out across a team is buying individual subscriptions, with the procurement and admin story that implies; the page says business account support is being worked on, not that it exists. This is the line that most often decides it for a CTO, and it has nothing to do with the agent's quality.
Publishing is a click. For a task started in the web app, you publish the branch or pull request yourself when you are satisfied with the diff; the label path posts a pull request link on the issue. That is a reasonable default for a tool that assumes a person is reviewing, and it is one more step than "come back to a PR."
Who should use Jules
A solo developer or a small team on GitHub, happy with Gemini, who wants to delegate the boring tickets and pay nothing. That is a large group of people and Jules serves them better than anything else on the market. The plan-approval step and the review behavior make it a safe first agent.
Look elsewhere if the backlog is in Linear, if you want to choose the model, or if you need to buy it for a team rather than a person. Sinatra is one of the alternatives for the first two: the GitHub half is the same gesture, a label on the issue, and it adds a Linear assignment as an equal trigger and a model credential you hold, pinned per repository. You give up the free allowance: our free tier is 5 tasks a day on your own key or subscription against Jules' 15. If the quota is why you are here, Jules is the right answer and the comparison page says so. For the GitHub half on its own, how to automatically fix GitHub issues with an agent covers the label workflow both tools share.
Common questions
Is Jules free?
Yes, with a published allowance of 15 tasks a day and 3 running at once. The higher tiers, 100 and 300 tasks a day, come with Google AI Pro and Ultra subscriptions.
Does Jules work with Linear?
As a tool, yes: MCP support lets a Jules task read and update Linear while it runs. As a trigger, no. There is no way to assign a Linear issue to Jules; the only documented tracker trigger is the jules label on a GitHub issue.
Can Jules use Claude or my own API key?
No. Jules runs Gemini models and documents no bring-your-own-key path.
Does Jules act on pull request review comments?
By default, yes. It reacts with 👀 to each comment it has read and pushes a commit with the requested changes. Reactive Mode limits it to comments that mention @Jules.
If you are choosing between Jules and an agent that also takes Linear tickets and your own model credential, the fair test is the same handful of issues through both, chosen before you know which tool gets them. Jules' side costs nothing. Sinatra's side costs nothing either on the free tier, with the tokens on the key or subscription you connect.