Kilo Code

OSS AI coding assistant for planning, building and fixing code

Vibe Coding

https://kilo.ai/

Kilo Code is an open-source AI programming assistant based on VS Code and JetBrains platform. It supports more than 500 AI models. With its model-agnostic positioning and zero markup strategy, Kilo Code has received more than 420,000 downloads within 6 months of launch, becoming an important player in the field of open-source AI programming tools. The core advantage of this product lies in its highly open model selection and Apache-2.0 open-source license, but it also faces room for improvement in user experience details (such as tab management, focus jump).

In-depth Report

Kilo Code is an open source AI programming assistant that runs on the VS Code and JetBrains platforms and supports more than 500 AI models. With its model-agnostic positioning and zero markup strategy, Kilo Code has received over 420,000 downloads within 6 months of launch, making it a major player in the open source AI programming tools space. The core advantage of this product is its highly open model selection and Apache-2.0 open source license, but it also faces room for improvement in user experience details (such as tab management, focus jumps).
Developed by the Kilo team, Kilo Code is an open source AI code assistant first released 6 months ago. The product won the title of "Open Source Product of the Month", has more than 2.3 million cumulative users, and has processed more than 25 trillion tokens. Kilo Code is forked based on Roo Code, innovatively adopts a multi-agent architecture, and is integrated into VS Code, JetBrains series IDEs (IntelliJ IDEA, PyCharm, WebStorm) and CLI command line environments. From the perspective of company background, the Kilo team has gained the trust of well-known technology companies such as Meta, Amazon, Airbnb, PayPal, Square, Red Hat, Grafana Labs, etc. Although specific financing information has not been disclosed in public materials, the rapid iteration of the product and community activity indicate that it has stable financial support. In terms of product positioning, Kilo Code clearly displays the banner of "model-agnostic" (model-agnostic), forming differentiated competition with closed AI programming tools such as Cursor and GitHub Copilot. Users can freely choose to use cutting-edge business models, bring their own API keys (BYOK mode), or deploy local models through tools such as Ollama and LM Studio to achieve true vendor neutrality.
Kilo Code provides six core modes, covering the entire process of software development. Code mode is used for writing, refactoring, and debugging production-level code, and is the most commonly used basic function. Architect mode focuses on architectural planning and provides structured guidance and technical solution design for complex functions. Debug mode has powerful debugging capabilities and can identify and fix bugs, track root causes of problems, and read execution traces to make targeted editing suggestions. Ask mode serves as a Q&A assistant to quickly answer code-related questions. Custom mode allows users to personalize the configuration and create exclusive workflows that meet the needs of the team. In addition, the Cloud Agents and Code Reviewer functions implement cloud agent services and automated code review capabilities. At the technical implementation level, Kilo Code is driven by OpenClaw (an open source AI agent framework) and provides access to more than 500 AI models through Kilo Gateway. The current offering offers free access to cutting-edge models like the Trinity Large Preview and MiniMax M2.5. In terms of integration, it supports mainstream communication tools such as Slack, Telegram, and Discord, making it convenient for users to manage AI agents through the chat interface. Judging from user experience feedback, the advantages of Kilo Code are reflected in multiple dimensions. The first is the deep integration with VS Code. Existing shortcut keys, terminals, and file navigation can continue to be used without having to re-adapt to the new IDE environment. Second is true model freedom, users can bring their own API keys and avoid vendor lock-in. The open source foundation also ensures the openness of the ecosystem, and users can participate in contributing code or deploy modifications by themselves. For teams with strict data policies, native model support provides privacy protection options. However, there are also obvious pain points in terms of user experience. The most obvious problem is the lack of tab functionality in the Kilo Code panel, which makes it impossible to handle multiple tasks or conversations at the same time. When working in parallel, you need to rely on the main editor tab. Another frequent pain point is forced focus switching when creating a new file, interrupting the workflow of reading or editing. In addition, although the multi-agent architecture is powerful, it takes one to two days for new users to understand various prompt words, guardrail settings and intervention timing, and the learning curve is relatively steep.
Kilo Code adopts a unique free open source + pay-as-you-go model. The extension itself is completely free and can be installed for free in the VS Code or JetBrains plugin marketplaces. The only cost is the model usage fee – users are responsible for paying for the API calls for their chosen AI model. The advantage of this model is that it is highly transparent and flexible. Users can bring their own API keys (BYOK mode), fully control the cost structure, and avoid being hijacked by unexpected subscription bills. The pricing of each model provider varies greatly, and users can choose the most cost-effective solution based on task requirements. For example, using a local model (via Ollama) can achieve zero marginal cost, while using cutting-edge business models (such as Claude, GPT-4) pay according to actual token consumption. For enterprise users, the Kilo team also provides a paid product KiloClaw, which is a managed AI agent that supports 24/7 operation and can be managed through channels such as Slack, Telegram, Discord, etc. Specific pricing is not displayed on the public page and may be for enterprise customers with higher SLA requirements. Compared with competing products, Cursor and GitHub Copilot adopt a closed source + subscription model, and the monthly fee is usually in the range of US$10-20. Kilo Code's free open source strategy significantly lowers the barriers to use, making it especially suitable for individual developers, small teams, and projects with limited budgets.
Judging from publicly available user reviews, Kilo Code has received diverse feedback. Positive comments focus on the following aspects. Many users speak highly of Kilo Code's freedom of model selection, believing this to be its core advantage over Cursor and Copilot. A user from Zhihu said, "Kilo Code was born: it perfectly integrates all the advantages of Cline and Roo Code, completely solves stuck bugs, supports 5 intelligent modes, has a free credit of 20 US dollars, automatically triggers context compression, intelligent task decomposition, real-time code interpretation, and doubles programming efficiency." Technical blogger Prathamesh pointed out in a detailed review that Kilo Code's deep integration in VS Code is the main advantage. "Existing shortcut keys, terminals, and file navigation can continue to be used without learning a new IDE." The experience feedback from the Chinese community is also positive. Some users shared their experience of using Kilo Code to develop web applications on social platforms, saying, "I used Kilo Code to develop a small website and debate on AI. It was quite interesting." Several tutorial articles on CSDN introduce the installation and practical application of Kilo Code in detail, and the overall evaluation is positive. Negative feedback mainly focuses on product details and experience. The review article mentioned above pointed out two main pain points: first, the lack of tab function and the inability to handle multiple tasks at the same time; second, automatic jump focus when creating a new file, interrupting the workflow. Although these problems do not affect core functions, they will affect efficiency in high-frequency usage scenarios. When discussing the comparison between Kilo Code and Cursor on Reddit, a user mentioned that Kilo's orchestration function is powerful but requires a certain learning cost. "New users need a day or two to understand the prompt words, guardrails, and intervention timing." Kilo Code has recently been on the hot list many times on Product Hunt, including products such as "Kilo Code for JetBrains" (670+ upvotes), demonstrating high community attention.
Judging from feedback from industry media and professional communities, Kilo Code has received high attention and positive reviews. At the technical community level, Kilo Code triggered discussions on Hacker News. Some developers shared their experience of using Kilo Code for "speedrun" (rapid development), believing that its multi-agent architecture has advantages in decomposing complex tasks. User reviews on open source community platforms such as SourceForge and Slashdot are generally positive, with most users recognizing their open source attributes and model flexibility. Reports from Chinese technology media also showed a positive attitude. Although there are no direct reviews on platforms such as 36Kr and Aifaner, technical articles in communities such as Zhihu and CSDN have introduced and recommended Kilo Code in detail. Some articles position Kilo Code as a tool that "subverts the programming experience" and believes that its multi-agent architecture is the core innovation that distinguishes it from traditional code completion tools. In terms of industry impact, Kilo Code’s open source strategy has had a catfish effect on the AI programming assistant market. Its "model-agnostic" positioning drives the industry toward more open model selection. Interestingly, when the Kuaishou technical team developed its own programming assistant KAT-Coder-Pro, it also used Kilo Code’s testing methods for actual testing, which reflects Kilo Code’s technical influence in the industry. In terms of competitive product landscape, Kilo Code's main competitors include Cursor (the benchmark for AI programming assistants), GitHub Copilot (Microsoft-based), Cline (open source competing products), Roo Code (upstream projects), etc. Compared with these competing products, Kilo Code's differentiated advantages are that it is completely open source, has zero price increase, and is model-independent; its disadvantage is that its brand awareness and maturity are relatively low, and some user experience details need to be polished.
As a rapidly iterating open source project, Kilo Code also faces some controversies and potential risks. Stability issues: Some users have reported that they have encountered "stuck" situations during use. Although subsequent versions have been improved, unresponsiveness may still occur when executing complex tasks. This may have an impact on complex projects that require long running times. Security considerations: Since Kilo Code can access the code repository and perform various operations, users need to pay attention to potential code leakage risks. It is recommended to prioritize using local models when working on sensitive projects, or carefully review AI-generated code. Maintenance risk: As an open source project, the long-term development of Kilo Code relies on community activity and team investment. If the core team reduces investment or undergoes personnel changes, it may affect continued product updates and technical support. Learning cost: Although the multi-agent architecture is powerful, the high learning curve may become an adoption barrier for users who pursue "out-of-the-box" functionality.
Kilo Code's target user group has obvious characteristics. The following are specific usage suggestions. Recommended installation scenarios: If you are a developer working in a VS Code or JetBrains environment, value the freedom of model selection, need to use local models in sensitive projects to meet data compliance requirements, or want to avoid vendor lock-in and manage API costs yourself, Kilo Code is a good choice. For individual developers or small teams with limited budgets, its free open source strategy can significantly lower the threshold for AI-assisted programming. Not a good fit: If you prefer an all-in-one, out-of-the-box experience, don’t want to spend time configuring and selecting models, or need a commercial product with full enterprise support and SLA, then Cursor or GitHub Copilot may be more suitable for your needs. Usage suggestions: Novice users are advised to start with official documents and tutorials, focusing on understanding the applicable scenarios of the six modes. Advanced users can use Custom mode to create team-specific workflows. Cost-sensitive users can give priority to trying local models (such as deploying domestic models such as Qwen and DeepSeek through Ollama) to achieve zero marginal cost.
As an important player in the field of open source AI programming assistants, Kilo Code provides developers and teams with a new AI-assisted programming option with its model-independent positioning, transparent pricing with zero markup, and Apache-2.0 open source license. Its support for 500+ models, broad compatibility with integrated VS Code/JetBrains/CLI, and complex task processing capabilities brought by its multi-agent architecture make it capable of competing with closed-source competitors such as Cursor in terms of technical capabilities. However, the product still has room for improvement in terms of user experience details (such as tab management, focus jump control), and the relatively steep learning curve may also affect the adoption of novice users. For users who pursue model freedom, open source ecology and cost transparency, Kilo Code is worth trying; for users who prefer integrated experience and mature commercial support, closed source competing products may be a safer choice. As the AI programming assistant market continues to evolve, whether Kilo Code's open source strategy and community-driven model can continue to lead industry innovation deserves continued attention.

User Reviews

Madison.ChavezK

—

今天再打开，已经是 OpenRouter 用量第一的开源 Agent 了，3M+ 用户。看着它从 Cline 分支一路长成这样，挺感慨。

DJohnson_66

—

Reddit 上大家都说它是「开源免费的 Cursor」，我觉得挺贴切。500+ 模型加 Ollama 本地跑，多模式工作流是真的香。社区也活跃，本地模型集成就是被用户反馈推着做的。要说缺点，就是 Zed、Neovim 还排不上一等公民，纯终端党可能还是更爱 Aider。

BlkckMax

—

新手得预留两三个小时折腾 API key 和选模型，不像 Copilot 开箱即用。但老手写起来很快就回不去了，尤其是要按任务切便宜模型和强模型的时候，成本能压下来一大截。

JacquelineBailey_X

—

免费装，模型费按原价付，穷人友好。

EFosterII

—

GitLab 联创搞的项目确实靠谱，800 万美元种子轮，三十来人远程团队，不是周末 side project。Apache 2.0 扩展完全可审计，能看见每一条发往模型的 prompt，没有偷偷压缩上下文或换模型。对要审计的团队来说这点值钱。

Billy297

—

Provider 路由三月那个 bug 真修好了，我一次会话里切了五次 Opus、GPT、DeepSeek、Gemini，账单都落在该落的地方。之前这毛病挺劝退的，现在终于稳了。

BenjaminBrooks007

—

比 Cursor 自由，但没它顺手，看你要啥。

happyelephant829

—

MCP 市场是隐藏宝藏，接了数据库和 Jira 之后，Kilo 能直接查工单、跑 SQL 再改代码，一条龙。不过功能铺得太开，有些模块还在成熟中，长会话偶尔有上下文漂移，复杂任务记得常开新会话。

PihlaWiita

—

CLI 二月就出了，终端里跑 headless 任务挺方便，但目前 Orchestrator 在 CLI 还不完整，CI 里跑自动化还是得靠 Claude Code。要纯 CI 头的别抱太大期望。

SusanKimQ889

—

本地 Ollama 跑模型，代码不出本机，安心。

刘梅萍

—

JetBrains 支持是我留下来的关键。早上 IntelliJ 写 Java，下午 VS Code 改前端，两边的 Agent 配置和逻辑一致，不用维护两套工具。WebStorm 里 Orchestrator 也能用，这点 Cline 做不到。团队里跨 IDE 的同事终于不用吵架了。

Patricia.RodriguezQ59

—

Architect 和 Orchestrator 别在同一工作区来回切，状态是共享的，我遇到过半截 Architect 计划漏进 Orchestrator 导致子代理把迁移跑了两遍。一个工作区用一种模式最稳。

Thomas_Ortiz_2020

—

行内补全还是比 Cursor 弱，复杂编辑才上 Kilo。

BrianMorris_Max314

—

在 22k 行的 Next.js 项目里实测了 0.13.2，多文件重构和按错误日志修 bug 都很顺。Orchestrator 把一个 Drizzle 迁移子任务丢给 Code 子代理，同时让另一个研究子代理去查 Postgres 文档，省下的上下文窗口很实在。代价是比直接跑 Claude Code 慢一点，wall-clock 大概 1.4 倍。

SBrooksIII40

—

Kilo Pass Starter 的积分不会跨月滚存，我五月初切回 BYOK 忘用了，白白过期了 11 刀额度。想用套餐的话记得当月花完，不然还是纯 BYOK 划算。

ThomasHill_2023

—

用了两周下来，最打动我的是零差价 BYOK。我自己对照 Anthropic 后台和 Kilo 的用量日志，账单精确到分对得上，没有偷偷加价。对于要控制成本的小团队来说这点太重要了，比 Cursor 那种订阅制透明多了。

dECLAN236

—

Orchestrator 模式确实好用，把一个全栈功能拆成 API、前端、测试三个子任务并行跑，省了不少来回切换的功夫。就是第一次配置模型路由花了点时间。

CarterKrause

—

刚升级到最新版 Kilo Code，Agent Manager 终于来了！可以同时跑多个代理任务，并行处理效率直接翻倍，这波更新太香了。

谭秀敏

—

从 Cline 切到 Kilo 了，开源真香。

石昊

—

实测多模型比较功能，选模型不再纠结，量化对比一目了然，开发者福音。

JCollins_Plus

—

基于 Kilo CLI 重构后执行速度快了很多，之前偶尔卡顿的问题基本解决了。

DRoberts39

—

内联代码审查功能上线了，行的级别直接注释，再也不用跳转到审查面板了，效率拉满。

lv924bbd

—

子代理功能太猛了，主代理调度子任务并行执行，大项目开发时间缩短将近一半。

VictoriaMiller_7

—

迁移到新版本后，任务时间线（Context Progress Graph）终于回来了，可视化追踪任务进度很实用。

GaëlLeclerc

—

新版 per-tool 权限系统比旧版 auto-confirm 细粒度多了，可以精确控制每个工具的执行权限，安全性提升明显。

yol2p2m

—

Modes 改名为 Agents 了，一开始有点不习惯，但用起来确实更直观，功能也更清晰了。

ZAkim

—

刚试了并行工具调用，同时执行多个操作的感觉太爽了，这才是真正的 AI 编程体验。

NSanders4

—

Profile 层简化成星级收藏了，虽然功能少了但界面清爽很多，点赞。

春雨522

—

Checkpoints 更名成 Snapshots 了，总觉得哪里怪怪的，不过功能没变就行。

彭兰

—

用 Kilo Code + Claude Sonnet 4 搭了一个全栈应用，前后端 + 数据库只花了 5 美元，太离谱了。

Kilo Code

In-depth Report

User Reviews

Similar Products