← Coding

JulesJules

Google

编程Coding 编码agentCoding agent 云端Cloud 厂商云Vendor cloud 常驻Always-on 审批Approval 订阅制Subscription 免费档Free tier 并行Parallel partial lifecycle: active 核验Checked 2026-10-23

本机什么都不装，任务跑在 Google 的云机器上，做完给你一个 PR。

它和你熟悉的 IDE / CLI 助手最大的差别在交互：任务异步执行，不实时对话。你提交任务、批准计划，然后走开，回头收 PR。

官方 GA 发布页原文："clones your codebase into a secure Google Cloud virtual machine (VM)"，以及 "operates asynchronously"。

Google's asynchronous coding agent: point it at a GitHub repo, and it clones the code into a fresh cloud VM, plans the change, and opens a pull request for review. The difference from the IDE/CLI agents on this site is that the interaction is not live — you approve the plan and walk away. The official REST docs make the default explicit: "By default, sessions created through the API will have their plans automatically approved. If you want to create a session that requires explicit plan approval, set the requirePlanApproval field to true"; and "The automationMode field is optional. By default, no PR will be automatically created." Tiers are bundled, not sold separately: the GA post describes a free "Introductory access to Jules," "Jules in Google AI Pro: 5x higher limits" and "Jules in Google AI Ultra: 20x higher limits," noting that "Specific usage limits are listed at jules.google."

适合与不适合When to use it, when not to

适合自包含、可验证的活：修一个明确 bug、升级依赖、补测试、迁移代码片段、小范围新功能。

它的产出是一个 PR，不是给你一个能对话的界面。

官方对交付形态的说法："test verified patches" 与 "Open PRs with runnable code and test results"

不适合两类活。一是需要实时 shell、边跑边改的高频交互，那是 CLI / IDE 形态的强项。二是需求模糊、需要反复澄清的架构级任务。官方 FAQ 自己就把「vague prompts」列进了任务失败的常见原因。

Self-contained, verifiable work: fixing a named bug, bumping dependencies, adding tests, migrating code snippets, small features — officially delivered as "test verified patches" and "PRs with runnable code and test results," which is exactly what makes it good for "submit the task and come back to a PR."

Not a fit for work that needs a live shell you can steer while it runs (CLI/IDE agents do that better), or vague, architecture-level tasks that need heavy clarification — the official FAQ itself lists "vague prompts" among the common causes of a failed task.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 直连不可达direct blocked　jules.google 大陆直连稳定超时（同日两次 000）。jules.google times out consistently from mainland (000 twice).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 模型由厂商指定，你选不了。

官方可读页里唯一点名模型的地方，是 GA 发布页。它说 Jules 用 Gemini 2.5 Pro 的进阶推理能力。Google One 套餐页在 Ultra 档写 Jules「access to latest models」，意思是更高档位能用更新的模型。

哪一档对应哪个模型，官方没写。

（已查 developers.google.com/jules/api 与 Google One AI 套餐页。）

官方原文：「Jules now uses the advanced thinking capabilities of Gemini 2.5 Pro」

这跟 IDE / CLI 那两类正好相反。它们通常有 /model 之类的选择面，而厂商云形态把「选模型」这件事整个收走了。
The model is vendor-assigned; you cannot choose it.

The only place the readable official pages name a model is the GA launch page. It says Jules uses the advanced reasoning capabilities of Gemini 2.5 Pro. The Google One plan page writes, for the Ultra tier, Jules "access to latest models", meaning higher tiers get newer models.

Which tier maps to which model is not written officially.

(developers.google.com/jules/api and the Google One AI plans page checked.)

Official wording: "Jules now uses the advanced thinking capabilities of Gemini 2.5 Pro"

This is the exact opposite of the IDE / CLI classes. Those usually offer a selection surface like /model, while the vendor-cloud form takes the whole "pick a model" decision away.
运行位置Where it runs: 任务跑在 Google 的云虚拟机上，本机不装任何东西。

每个任务一台全新的云 VM：在里面 clone 你的仓库、装依赖、改代码。

官方 FAQ 原文：「Each task runs in a fresh virtual machine where Jules clones your repo, installs dependencies, and makes changes based on your prompt」 官方 GA 发布页把这条列为卖点：「Tasks run inside a cloud VM, enabling concurrent execution」

除了网页（jules.google），还有两个入口：Jules REST API （jules.googleapis.com/v1alpha）和 Jules Tools CLI。 但执行永远在厂商云上，客户端只负责下达和查看。
Tasks run on Google's cloud VMs; nothing is installed locally.

Each task gets a brand-new cloud VM: inside it Jules clones your repo, installs dependencies, and changes code.

Official FAQ wording: "Each task runs in a fresh virtual machine where Jules clones your repo, installs dependencies, and makes changes based on your prompt" The official GA launch page lists this as a selling point: "Tasks run inside a cloud VM, enabling concurrent execution"

Besides the web app (jules.google), there are two more entry points: the Jules REST API (jules.googleapis.com/v1alpha) and the Jules Tools CLI. But execution always happens on the vendor cloud; clients only submit and observe.
本地文件Local files: 默认跟你本机完全隔离。

Jules 只在云 VM 里 clone 仓库干活，不碰你本机的文件。这个 VM 有联网能力 —— 这既是它的能力，也是它的风险面。

官方 FAQ：该 VM「executed in a secure, cloud-based virtual machine (VM) with internet access」，并要求用户「treat the environment with the same security precautions you would for any public or shared compute surface」 官方同时提醒：不要把密钥、token 之类提交进仓库。

哪些仓库对 Jules 可见，由 Jules GitHub app 的授权范围决定（FAQ 指引到 GitHub 的 applications 设置里逐项调整）。
By default it is fully isolated from your local machine.

Jules only clones repos and works inside the cloud VM; it never touches files on your machine. This VM has internet access — which is both a capability and a risk surface.

Official FAQ: the VM is "executed in a secure, cloud-based virtual machine (VM) with internet access", and users are told to "treat the environment with the same security precautions you would for any public or shared compute surface". The official pages also warn: do not commit secrets, tokens and the like into the repo.

Which repos are visible to Jules is decided by the authorization scope of the Jules GitHub app (the FAQ points to GitHub's applications settings for per-repo adjustment).
关机后的任务Tasks after shutdown: 官方将后台任务描述为异步执行。

Jules operates asynchronously, allowing you to focus on other tasks while it works in the background

You can leave the app after submitting a task

只需打开通知。任务失败会自动重试。

Jules will retry automatically

2025-12-10 的官方更新再补两项主动能力：

Suggested Tasks —— 在最多五个仓库上主动扫描代码、提出改进，先支持 #todos
Scheduled Tasks —— 按你设定的节奏自动执行，如依赖检查、每周例行

Scheduled Tasks ... you can now define the cadence and Jules will perform the task at that time

并发任务数随订阅档位变化（具体数字见官方套餐页，本站按本站的报价记录约定不存）。
This is its thickest dimension; the official pages define it as "asynchronous" throughout.

Jules operates asynchronously, allowing you to focus on other tasks while it works in the background

You can leave the app after submitting a task

Just turn on notifications. Failed tasks retry automatically.

Jules will retry automatically

The official 2025-12-10 update adds two proactive capabilities:

Suggested Tasks — proactively scans code and proposes improvements across up to five repos, starting with #todos
Scheduled Tasks — runs automatically at your defined cadence, e.g. dependency checks, weekly routines

Scheduled Tasks ... you can now define the cadence and Jules will perform the task at that time

Concurrent task counts vary by subscription tier (concrete numbers live on the official plan page; this site does not store them per the price-recording policy).
工具与扩展Tools & extensions: 工具面集中在云 VM 内的执行能力 + GitHub 集成。

官方 FAQ 与 GA 页确认了 VM 内的执行能力。包括装依赖、跑测试、按 setup 脚本构建。支持的语言是 JavaScript/TypeScript、Python、Go、Java、Rust。

language agnostic but works best with

（实际取决于 VM 里装了什么。）

不支持长驻进程 —— npm run dev 之类跑不起来。

Long-running processes like dev servers or watch scripts aren't currently supported in setup scripts

另有前端验证能力：在 VM 内跑脚本、回传截图。见官方发布页的 web-app 测试说明。

Jules 是否支持接入 MCP server：官方未说明。 （已查 developers.google.com/jules/api 与官方 FAQ 页。）
The tool surface centers on execution capability inside the cloud VM plus GitHub integration.

The official FAQ and GA page confirm the VM's execution capabilities: installing dependencies, running tests, building via setup scripts. Supported languages are JavaScript/TypeScript, Python, Go, Java, Rust.

language agnostic but works best with

(in practice it depends on what is installed in the VM.)

Long-running processes are not supported — things like npm run dev will not run.

Long-running processes like dev servers or watch scripts aren't currently supported in setup scripts

There is also front-end verification: run scripts inside the VM and get screenshots back. See the web-app testing notes on the official launch page.

Whether Jules supports connecting MCP servers: not stated officially.

(developers.google.com/jules/api and the official FAQ page checked.)
上下文与记忆Context & memory: 以 Session / Activity 为上下文单位。

Session — A continuous unit of work within a specific context, similar to a chat session. A session is initiated with a prompt and a source

Activity — A single unit of work within a Session

一个 Session 含用户与 agent 双方的多条 activity。比如生成计划、发消息、更新进度。

也就是说面向程序化调用时，上下文是「会话 + 活动」这条可轮询的状态流，而不是一次性的对话。跨任务层面，官方提供环境快照复用：

reusing previous setups so new tasks run faster

上下文窗口大小与长期记忆的具体实现，官方可读页未给数字 —— 官方未说明。 （已查 Jules API 页与官方 FAQ 页。）
Session / Activity is the unit of context.

Session — A continuous unit of work within a specific context, similar to a chat session. A session is initiated with a prompt and a source

Activity — A single unit of work within a Session

One Session contains multiple activities from both the user and the agent, e.g. plan generation, messages, progress updates.

In other words, for programmatic use the context is the pollable "session + activity" state stream, not a one-shot conversation. Across tasks, the official pages offer environment snapshot reuse:

reusing previous setups so new tasks run faster

Context-window size and the concrete implementation of long-term memory — no numbers given on the readable official pages; not stated officially.

(Jules API page and official FAQ page checked.)
权限与限制Permissions & limits: 审批的核心是「计划审批」，且 API 与网页默认值不同。

By default, sessions created through the API will have their plans automatically approved. If you want to create a session that requires explicit plan approval, set the requirePlanApproval field to true

网页端则默认：

官方原文："Prompt a task and approve Jules' plan"

即程序化入口默认跳过人工审批，需显式打开。

代码落地方式由创建 session 时的 automationMode 决定 —— 设 AUTO_CREATE_PR 才会自动开 PR。

The automationMode field is optional. By default, no PR will be automatically created

仓库访问走 Jules GitHub app 的授权范围。

隐私口径：官方 FAQ 明确它不用私有仓库内容训练。

Jules does not train on private repository content

额度边界：免费层与 Pro / Ultra 的具体任务数，官方可读页未列（见官方套餐页）。但「额度按任务计、不按 token 计」这一点，从官方 API 概念（session 即任务单位）可以读出来。
The core of approval is "plan approval", and the API and web defaults differ.

By default, sessions created through the API will have their plans automatically approved. If you want to create a session that requires explicit plan approval, set the requirePlanApproval field to true

The web default is:

Official wording: "Prompt a task and approve Jules' plan"

i.e. the programmatic entry skips human approval by default; it must be explicitly turned on.

How code lands is decided by the automationMode set when creating the session — only AUTO_CREATE_PR opens a PR automatically.

The automationMode field is optional. By default, no PR will be automatically created

Repo access follows the authorization scope of the Jules GitHub app.

Privacy stance: the official FAQ states clearly it does not train on private repo content.

Jules does not train on private repository content

Quota boundary: the concrete task counts for the free tier and Pro / Ultra are not listed on the readable official pages (see the official plan page). But the point that "quota is counted per task, not per token" can be read from the official API concept (session is the task unit).
适合什么任务Best for: 官方给自己划的活是具体的：把任务描述交进 GitHub 仓库后，它做这些事：

官方原文："fix bugs, update dependencies, migrate code, and add new features"
交付方式 官方原文："test verified patches"、"Open PRs with runnable code and test results"

因此适合自包含、可验证的活：修一个明确的 bug、升依赖版本、补测试、迁移代码片段、小范围新功能。 不适合：需要实时 shell、边跑边改的高频交互。这类活本分区 CLI / IDE 形态更合适。也不适合需求模糊、需要大量澄清的架构级任务。官方 FAQ 自己把「vague prompts」列为任务失败的常见原因之一。
The work the official pages scope for it is concrete: hand a task description to a GitHub repo and it does these things:

Official wording: "fix bugs, update dependencies, migrate code, and add new features"
Delivery form official wording: "test verified patches", "Open PRs with runnable code and test results"

So it fits self-contained, verifiable work: fixing a clear bug, bumping a dependency, adding tests, migrating code snippets, small-scope new features. It does not fit: real-time shell work and high-frequency change-as-it-runs interaction — for that, the CLI / IDE forms in this section are better. It also does not fit vague, architecture-level tasks needing heavy clarification. The official FAQ itself lists "vague prompts" among common causes of task failure.

头号误解Top misconception

把 beta 旧文档的「2 并发 / 每天 5 次」当成现行免费额度Treating the old beta docs' "2 concurrent / 5 total tasks per day" as today's free tier
以为 API 建的任务都会等你批准计划，官方默认自动批准Assuming API-created sessions wait for you to approve the plan — they auto-approve by default
把 Jules 当成本地 / IDE 里的实时 agent——它是异步云 agentMistaking Jules for a local/IDE live agent — it is an asynchronous cloud agent

未知项清单Known unknowns

Jules 当前默认模型与各档位的模型映射——官方未说明（已查 developers.google.com/jules/api 与 Google One AI 套餐页）
逐档每日任务数与并发数——官方未说明（已查 jules.google 定价页、Google One AI 套餐页与 GA 发布页）
Jules 是否支持接入 MCP server——官方未说明（已查 Jules API 页与官方 FAQ 页）
上下文窗口大小与长期记忆的实现
云 VM 的单任务时长与并发上限
是否支持 GitHub 之外的代码托管

Jules's current default model and the per-tier model mapping — not stated officially (developers.google.com/jules/api and the Google One AI plans page checked)
Per-tier daily task counts and concurrency — not stated officially (jules.google pricing page, Google One AI plans page and the GA launch page checked)
Whether Jules supports connecting MCP servers — not stated officially (Jules API page and official FAQ page checked)
Context-window size and the implementation of long-term memory
Per-task duration and concurrency ceiling of the cloud VM
Whether code hosting beyond GitHub is supported

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
changelog	Google · Jules 官方变更记录（2026-10-08 核验，官方 changelog 入口）Google · Jules official changelog (verified 2026-10-08, official changelog entry)	https://jules.google/docs/changelog/
docs	Google · Jules API（alpha，Sources / Sessions / Activities、X-Goog-Api-Key、计划审批与 AUTO_CREATE_PR）Google · Jules API (alpha; Sources / Sessions / Activities, X-Goog-Api-Key, plan approval and AUTO_CREATE_PR)	https://developers.google.com/jules/api
docs	Google · Jules REST 参考（approvePlan / create / list / sendMessage 等方法）Google · Jules REST reference (approvePlan / create / list / sendMessage and other methods)	https://developers.google.com/jules/api/reference/rest
docs	Google · Jules 官方文档 FAQ / 用量限制（云 VM、联网、失败重试、不训练私有仓库）Google · Jules official docs FAQ / usage limits (cloud VM, internet access, failure retry, no training on private repos)	https://jules-documentation.web.app/faq
pricing	Google · Google AI 套餐页（Pro 列 Jules 扩展额度、Ultra 列任务与并发扩展，2026-10-08 取）Google · Google AI plans page (Pro lists Jules extended quota, Ultra lists task and concurrency extensions, captured 2026-10-08)	https://one.google.com/intl/en_us/about/google-ai-plans/
engineering	Google · Jules 正式可用（2025-08-06 GA：免费入口与 Pro 5x / Ultra 20x 额度）Google · Jules generally available (2025-08-06 GA: free entry with Pro 5x / Ultra 20x quota)	https://blog.google/technology/google-labs/jules-now-available/
engineering	Google · Jules 主动能力更新（Suggested Tasks / Scheduled Tasks / Render 集成，2025-12-10）Google · Jules proactive-capability update (Suggested Tasks / Scheduled Tasks / Render integration, 2025-12-10)	https://blog.google/technology/developers/jules-proactive-updates/

实测记录Measured runs

本档未跑端到端编码任务；能力结论来自官方文档，编码向测量协议见 protocols/coding/_protocol.md。

实测建议（三条，都指向官方可读页没给的东西）：

计划审批开关的真实行为 —— API 默认自动批准，网页端要批准，两处默认值不一样
关掉浏览器后任务会不会继续，完成时是否真的开出 PR
免费层与 Pro 档的实际任务数差多少 —— 这个数官方页没写，只能实测确认

This site has not completed measured runs. The test protocol is at protocols/coding/_protocol.md.

Suggested measurement (three items, all pointing at what the readable official pages do not give):

The real behavior of the plan-approval switch — the API defaults to auto-approve, the web requires approval; the two defaults differ
Whether tasks keep running after you close the browser, and whether a PR really gets opened on completion
How much the free tier and Pro tier actually differ in task counts — that number is not on the official pages; only measurement can confirm it

异步流程与计划审批（核验重点）The async flow and plan approval (this verification's focus)

官方 REST 文档把流程写得很清楚，其中有一处最容易被误读：

官方原文：By default, sessions created through the API will have their plans 官方原文：automatically approved. If you want to create a session that requires 官方原文：explicit plan approval, set the requirePlanApproval field to true.

即：API 建的任务默认自动批准计划，要人工把关必须显式打开。 网页端默认则是「Prompt a task and approve Jules' plan」。

PR 的生成同样要显式开关：

官方原文：The automationMode field is optional. By default, no PR will be 官方原文：automatically created.

设成 AUTO_CREATE_PR 才会自动开 PR。

进度以 activity 流的形式返回。相关 API 方法和事件名：

方法：sessions/{id}:approvePlan、sessions、sessions/{id}:sendMessage 事件：planGenerated / progressUpdated / sessionCompleted 产物：bashOutput、changeSet（gitPatch）

The official REST docs spell the flow out clearly, and one point is the easiest to misread:

Official wording: By default, sessions created through the API will have their plans Official wording: automatically approved. If you want to create a session that requires Official wording: explicit plan approval, set the requirePlanApproval field to true.

i.e.: tasks created via the API have their plans auto-approved by default; human gating must be explicitly turned on. The web default is "Prompt a task and approve Jules' plan".

PR generation likewise needs an explicit switch:

Official wording: The automationMode field is optional. By default, no PR will be Official wording: automatically created.

Only AUTO_CREATE_PR opens PRs automatically.

Progress comes back as an activity stream. Related API methods and event names:

Methods: sessions/{id}:approvePlan, sessions, sessions/{id}:sendMessage Events: planGenerated / progressUpdated / sessionCompleted Artifacts: bashOutput, changeSet (gitPatch)

核验说明Verification notes

confidence: partial 的依据：已核到了官方 GA 发布页、Google One 套餐页、 Jules API 与其 REST 参考、官方 FAQ 页、两份官方博客更新。

缺口集中在两类，而且这两类在可读的官方页里确实没有：

「数字」：逐档每日任务数与并发数（jules.google 定价页是 JS 渲染的，取不到）
「模型映射」：当前默认模型、各档位对应哪个模型
以及：是否支持 MCP

这些在本档案中记为「官方未说明」，并列出所查页面 —— 不猜。

Basis for confidence: partial: the 2026-10-08 check verified the official GA launch page, the Google One plan page, the Jules API and its REST reference, the official FAQ page, and two official blog updates.

The gaps fall into two classes, and these two classes genuinely are absent from the readable official pages:

"Numbers": per-tier daily task counts and concurrency (the jules.google pricing page is JS-rendered and could not be captured)
"Model mapping": the current default model and which model each tier maps to
Also: whether MCP is supported

These are recorded as "not stated officially", with the checked pages listed — no guessing.