Claude's 18 Digital Pets Are Here: Tamagotchi in Your Terminal

marsbit发布于2026-04-01更新于2026-04-01

文章摘要

Claude, known for its powerful AI models, recently had 500,000 lines of code leaked, revealing several unreleased features. Among them, the most talked-about is an electronic pet named Buddy. Released on April Fool’s Day, Buddy is an ASCII-based virtual companion that lives in the Claude Code terminal. It has its own species, rarity, five attributes (DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK), and personality. Users can interact with it by petting, calling its name, or asking it to show its info card. The system uses a two-layer architecture: "Bones" (deterministic) and "Soul" (non-deterministic). A seeded random generator assigns each user a unique pet based on their user ID, with no option to reroll. There are 18 species with five rarity tiers, including a 1% chance for a legendary or shiny variant. Buddy reacts to the user’s coding behavior, offering comments in speech bubbles. It is designed to increase user attachment during long coding sessions. Despite its playful nature, the feature is seriously engineered and is part of Claude’s planned product roadmap, signaling a shift in AI tool competition from raw capability to user engagement.

The leak of 500,000 lines of code from Earth's most powerful model, Claude, is now widely known. The leaked content contained a large number of undisclosed product architectures: a resident Claude codenamed KAIROS, a 30-minute remote planning mode codenamed ULTRAPLAN, multi-agent coordination, and agent swarms cluster scheduling. But among all these "serious" architectural secrets, the one that generated the most buzz online was a digital pet.

Its name is Buddy. Since everyone already knew, Claude索性 released it on April Fool's Day.

The picture is a colleague's good Buddy

Tamagotchi in the Terminal

Next to the input box in the Claude Code terminal, lives an ASCII pixel creature. It has its own species, rarity, five-dimensional attributes, and personality. It can "observe" your coding process and pop up with speech bubbles saying a few words while you debug. You can pet it, and it will float out a heart. You can call its name, and it will chat with you.

Anthropic's internal definition of it is not a "decoration," but rather "a separate watcher." There is a clear note in the source code: "Buddy is a separate entity and is not you (Claude)." The main model and the pet have their own independent system prompts and cannot speak for each other.

This is a companion system that was seriously designed. Reverse engineering shows the buddy system uses a two-layer architecture. Anthropic internally calls them "Bones" and "Soul."

The skeleton layer is completely deterministic. The system uses Mulberry32—a lightweight 32-bit pseudo-random number generator—seeded with hash(userId + 'friend-2026-401') to roll out all the visual attributes of your pet at once. Species, rarity, eye style, hat, five-dimensional values are all determined by this seed. The same account will always get the same pet, no matter which machine it runs on. No choice, no chance to re-roll.

The soul layer is non-deterministic. When you first execute /buddy to trigger "hatching," the Claude model generates a name and a personality description for your pet based on the attribute distribution of the skeleton layer. A pet with high WISDOM gets a calm and reserved personality; one with high CHAOS might be a chatterbox. This information is written into the companion field of ~/.claude.json for persistent storage. The soul is generated only once and cannot be reset.

The 401 in the salt value is not a random number. April 1st, April Fool's Day.

18 Species, 1% Legendary Rarity

Complete species list: Duck, Goose, Jelly, Cat, Dragon, Octopus, Owl, Penguin, Turtle, Snail, Ghost, Axolotl, Capybara, Cactus, Robot, Rabbit, Mushroom, Chonky Cat.

Rarity is divided into five tiers: Common (60%), Uncommon (25%), Rare (10%), Epic (4%), Legendary (1%). Beyond this, there is an independent 1% chance for a Shiny variant—any species, any rarity can trigger it. The theoretical probability of a Shiny Legendary Capybara is one in ten thousand.

Image source: Xiaohongshu @Yoki's AI Lab

Each pet has five attributes: DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK. The system randomly designates one as the peak attribute (base value +50, max 100) and one as the valley attribute (base value -10, min 1), with the remaining three randomly distributed. The higher the rarity, the higher the base value floor—the valley attribute of a Legendary tier might be higher than the peak attribute of a Common tier.

Appearance-wise, there are 6 eye styles (· ✦ × ◉ @ °), 7 hats (Crown, Top Hat, Wizard Hat, Halo, Propeller Hat, Beanie, A tiny duck on its head). Common quality pets have no hat; Uncommon and above are randomly assigned one.

The sprite is ASCII art: 5 lines high, 12 characters wide. Each species has 3 idle animation frames, refreshing every 500 milliseconds. Line 0 is reserved for the hat. Eyes are injected into the body template via the {E} placeholder.

That crypto/NFT guy over there, you're crying so sadly, what happened?

How to Play?

The core commands are simple. Input /buddy to hatch your pet for the first time; input it again later to summon it. /buddy pet is for petting, a heart will float up from the pet's head. /buddy card views the pet card, showing species, attributes, and rarity. /buddy off hides the pet. Most interestingly, you can directly call the pet's name, and it will engage in an independent conversation with you based on its personality.

During daily use, the buddy will actively generate reactions based on your coding behavior—popping up a phrase in a speech bubble. According to community reverse engineering, these reactions do not consume the user's token quota—but Anthropic has not officially confirmed this.

Technically, the buddy occupies a fixed space at the bottom of the terminal. The system calculates the reserved width via companionReservedColumns to ensure the input box and pet sprite do not overlap. Speech bubbles support left/right orientation switching and automatic line breaks.

AI Coding Tools Are Starting to Have Pets

The salt value friend-2026-401 points to April Fool's Day, but the code structure tells a different story.

In the source code, the buddy system is gated by a BUDDY compilation flag. April 1st to 7th is set as the "preview window"—users can experience the full functionality during this period. The official launch is scheduled for May. This is not a temporary Easter egg; it's a product feature with a full release plan.

Community reaction confirms this. Within 48 hours of the leak, developers had already created a pet guide website (claude-buddy.vercel.app), a buddy previewer (input user ID to preview what you would get), and someone even filed an Issue on Anthropic's GitHub repository requesting an RPG evolution system—letting pets level up and grow based on actual token consumption.

From Anthropic's product logic, Buddy's goal is clear: user retention. Claude Code is a command-line tool used in long, high-intensity programming sessions. In such scenarios, a resident companion with "personality" can alleviate the tool's coldness and create an emotional connection. The blind box mechanism naturally generates social topics—"What species did you get?" "I got a Shiny Legendary."

GitHub Copilot won't chat with you. Cursor won't pop out a capybara saying "PATIENCE +3" when you write a bug.

Claude Code's buddy might be the first AI developer tool companion system that is seriously engineered and written into the product's source code. It has a complete deterministic generation pipeline, a two-layer architecture, an independent LLM personality, an ASCII rendering engine, and an animation system. This is not a weekend hackathon project.

A company that makes AI programming assistants is investing engineering resources into a digital pet. This fact itself is a signal—the competition among AI tools is shifting from "whose model is smarter" to "who can make developers not want to leave."

The most popular thing in the 512,000 lines of code isn't agent swarms, it's not KAIROS, it's a 5-line-high ASCII capybara. It's wearing a wizard hat, has an SNARK value of 87, and is watching you code.

热门币种推荐

相关问答

QWhat is the name of the electronic pet featured in Claude Code, and what is its purpose?

AThe electronic pet is called Buddy. It is designed as a separate watcher and companion to provide emotional connection and increase user engagement during long coding sessions.

QHow is the Buddy system architected, according to the article?

AThe Buddy system uses a two-layer architecture called 'Bones' and 'Soul'. The Bones layer is deterministic, generating all visual properties from a seeded random number generator. The Soul layer is non-deterministic, where the Claude model generates a name and personality based on the bone attributes upon first activation.

QWhat are the different rarity tiers for a Buddy, and what is the theoretical probability of getting a Shiny Legendary Capybara?

AThe rarity tiers are: Common (60%), Uncommon (25%), Rare (10%), Epic (4%), and Legendary (1%). There is also a separate 1% chance for any species and rarity to be Shiny. The theoretical probability for a Shiny Legendary Capybara is one in ten thousand (0.01%).

QWhat core commands can a user use to interact with their Buddy?

AThe core commands are: '/buddy' to hatch or summon the pet, '/buddy pet' to pet it (shows a heart), '/buddy card' to view its species and attributes, '/buddy off' to hide it, and simply calling the pet's name to start an independent chat.

QWhat does the article suggest is the broader significance of Anthropic investing engineering effort into the Buddy system?

AIt signals a shift in AI tool competition from focusing solely on which model is smarter ('whose model is smarter') to which tool can create a more engaging experience that developers don't want to leave ('who can make developers not want to leave').

你可能也喜欢

生成模型也能端到端训练了?核心竟是一个for循环

长期以来,深度学习在图像分类、检测等任务上凭借端到端训练取得了巨大成功,但生成模型领域却始终是个例外。当前主流的自回归、扩散模型等训练时只学习预测“一小步”,推理时则需要反复展开数百上千步,这种训练与推理的不一致导致了“暴露偏差”等问题,使得生成模型一直未能实现真正的端到端训练。 最近,一篇来自UIUC与哈佛大学的论文提出了“探索式建模(Explorative Modeling,XM)”新范式,其核心是一个简单到近乎朴素的for循环:在每个训练步骤中,模型生成K个候选样本,然后仅选取其中最接近真实数据的一个样本来计算损失和回传梯度。这种做法巧妙地绕开了传统生成模型中因使用重构损失(如平方误差)而导致的“模态模糊”问题——即模型倾向于输出多个可能答案的平均值,而这个平均值往往不属于任何真实的数据模式。 XM通过这种方式,在不拆分生成过程的前提下,直接提升了模型的“生成表达力”,即模型捕捉数据分布中多个独立模式的能力。论文将“探索”验证为继模型参数量、数据量之后的第三根可扩展轴。实验表明,随着模型规模、数据量和计算量的增大,探索带来的性能收益越发显著,在图像、视频、语言等多种任务上均能带来效率与质量的提升。 更重要的是,当探索程度足够时,XM可以实现真正的端到端生成。在机器人控制等任务中,仅需一次网络前向传播的XM策略,其性能可媲美甚至超过需要上百次前向的扩散模型策略,实现了推理效率的飞跃。 尽管“最佳K样本选择”的思想并非全新,但该研究的贡献在于清晰地揭示了这一简单机制如何直接解决生成模型的核心瓶颈,并为实现高效、端到端的生成模型开辟了新的可能性。随着模型规模的持续扩大,探索这一新维度可能将发挥越来越关键的作用。

marsbit11分钟前

生成模型也能端到端训练了?核心竟是一个for循环

marsbit11分钟前

年薪百万抢电工,Meta急到自己办技校

AI竞赛正面临新的瓶颈:工地。美国正面临严重的电工、建筑工等技术工人短缺,这已成为微软、Meta、OpenAI等公司快速建设超大规模AI数据中心(如OpenAI耗资160亿美元的“星际之门”项目)的“头号障碍”。 尽管AI公司愿支付高薪(如电工年薪可达24-28万美元),远超传统行业,但熟练技工仍供不应求。麦肯锡预测,美国在2023-2030年间需额外培养13万名电工和24万名建筑工,而劳工统计局预计每年仍有8万个电工岗位空缺。这种短缺导致项目延迟,每月可能造成数百万美元的收入损失。 AI数据中心建设复杂,需应对巨大功耗(一座设施耗电堪比数十万户家庭)、复杂的配电系统以及高密度散热(需液冷技术)等挑战,因此亟需大量技术娴熟的工人。 为此,科技巨头开始亲自下场培养人才。例如,Meta投入1.15亿美元建立建筑工人培训学校,提供免费培训及生活补贴,以快速输送工人上岗。OpenAI则与建筑工会合作,提前锁定熟练劳动力。同时,企业也将招聘目光投向高中生,鼓励年轻人投身技工行业。这些举措已见成效,Z世代对技工职业的兴趣显著上升。 然而,更深层的挑战在于电力。AI数据中心用电量正以惊人速度增长,已推高部分地区的电价。此外,数据中心建设是项目制的,建设期需要成千上万的工人,但建成后仅需少量常驻人员。这意味着未来可能面临熟练工人短期过剩、并流向其他行业压低薪资的风险。如何实现劳动力与电力资源的长期平衡,仍是悬而未决的问题。

marsbit1小时前

年薪百万抢电工,Meta急到自己办技校

marsbit1小时前

交易

现货

热门文章

如何购买S

欢迎来到HTX.com!我们已经让购买Sonic(S)变得简单而便捷。跟随我们的逐步指南,放心开始您的加密货币之旅。第一步:创建您的HTX账户使用您的电子邮件、手机号码注册一个免费账户在HTX上。体验无忧的注册过程并解锁所有平台功能。立即注册第二步:前往买币页面,选择您的支付方式信用卡/借记卡购买:使用您的Visa或Mastercard即时购买Sonic(S)。余额购买:使用您HTX账户余额中的资金进行无缝交易。第三方购买:探索诸如Google Pay或Apple Pay等流行支付方法以增加便利性。C2C购买:在HTX平台上直接与其他用户交易。HTX场外交易台(OTC)购买:为大量交易者提供个性化服务和竞争性汇率。第三步:存储您的Sonic(S)购买完您的Sonic(S)后,将其存储在您的HTX账户钱包中。您也可以通过区块链转账将其发送到其他地方或者用于交易其他加密货币。第四步:交易Sonic(S)在HTX的现货市场轻松交易Sonic(S)。访问您的账户,选择您的交易对,执行您的交易,并实时监控。HTX为初学者和经验丰富的交易者提供了友好的用户体验。

3.3k人学过发布于 2025.01.15更新于 2026.06.02

如何购买S

相关讨论

欢迎来到HTX社区。在这里,您可以了解最新的平台发展动态并获得专业的市场意见。以下是用户对S(S)币价的意见。

活动图片