AIWiki
Malaysia
Back to all articles
Tools & Platformskilo codeai coding agentopen source

Kilo Code

4 min readUpdated October 2026
Kilo Code
Type
Open-source AI coding agent
Developer
Kilo-Org (acquired by Anaconda, July 2026)
First released
March 2025
Licence
MIT
Platforms
VS Code, JetBrains, CLI, cloud, Slack
Key features
Orchestrator mode, parallel subagents, 500+ models, zero markup
Related
Cline, Roo Code, Cursor, GitHub Copilot
Kilo Code is a free and open-source artificial intelligence coding agent that reads and edits project files, runs terminal commands, drives a browser and debugs software from natural-language instructions. It is model-agnostic: the same agent can be pointed at Claude, GPT, Gemini, DeepSeek, Kimi or a locally hosted model without changing the tool itself.[1]

History

Kilo Code launched in March 2025 as a fork of Roo Code, which had itself been forked from the open-source coding agent Cline, and subsequently blended features from both projects while building its own agent harness and model-routing layer.[1] The company was founded by Scott Breitenother, Emilie Schario and Sid Sijbrandij, a co-founder of GitLab.[1]

In December 2025 the project raised an US$8 million seed round led by Cota Capital, with General Catalyst, Breakers, Quiet Capital and Tokyo Black participating. At the time of the announcement Kilo reported more than 750,000 downloads, first place on OpenRouter by coding token volume and over six trillion tokens processed per month.[2] The repository passed 27,000 GitHub stars by October 2026, and the company claimed more than three million developers using the agent across its surfaces.[3]

On 15 July 2026, Anaconda — the Python distribution and data-science platform company — acquired Kilo Code for an undisclosed sum, following its earlier acquisition of Outerbounds. The product, plans and open-source repository continued without announced changes after the deal.[4]

Key Concepts and Technology

Kilo Code's differentiator is Orchestrator mode: instead of one agent executing a long task in sequence, Orchestrator decomposes the work into subtasks and routes each to a specialist mode — Architect for planning, Code for implementation and Debug for diagnosis. Subagents can run in parallel across separate git worktrees, with inline diff review and either per-step approval or automatic approval for trusted workflows.[3]

The agent is model-agnostic by design. It connects to more than 500 models through its Kilo Gateway and OpenRouter, accepts direct provider API keys, can reuse existing subscription plans, and can fall back to locally hosted models through Ollama, LM Studio or OpenAI-compatible servers such as vLLM — an Auto Model feature routes individual tasks by task type, quality and cost.[1] Billing is bring-your-own-key at provider rates with zero markup, or managed credits through Kilo Pass from US$19 per month.[4]

Extensibility comes through Model Context Protocol (MCP) servers, custom agents defined in Markdown files and a skills format, plus GitHub integration for automated code review. The extension and CLI are open source under the MIT licence, so code stays on a team's own machines unless a cloud model is deliberately called.[3]

Applications and Impact

Kilo Code is used for feature development, refactoring, test generation, debugging and code review, and is positioned for engineering teams that want agentic coding without per-seat licensing — the extension is free and teams pay only for the tokens they consume. The project is among the most popular open-source coding agents and ranks first on OpenRouter by coding token volume.[2] Its rise, alongside Cline and the now-archived Roo Code, established a model-agnostic open-source tier of AI developer tooling alongside proprietary subscriptions such as Cursor and GitHub Copilot.

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: for Malaysian software houses, agencies and freelance developers the cost equation matters more than brand: a free MIT-licensed agent paired with low-cost open-weight models (DeepSeek, Kimi, Qwen, GLM) costs a fraction of per-seat Western subscriptions, which is why open-source bring-your-own-key agents spread quickly among local startups and solo builders. MDEC's digital-skills programmes and HRD Corp-funded training courses increasingly teach agentic coding workflows, and universities using VS Code in lab classes can adopt the tool without a procurement exercise.[5]

Two guardrails apply. Under the Personal Data Protection Act 2010, source code and repositories containing customer or employee data remain personal data where they identify individuals: sending them to a third-party model API is a cross-border transfer requiring a lawful basis, consent in the regulated cases, and vendor due diligence — a step Malaysian banks, insurers and healthcare providers already apply to outsourced IT. Second, because the agent can execute terminal commands, local enterprises should keep scoped permissions and human approval switched on before letting it touch production systems — a pattern the platform's own approval model supports.[5]

References

  1. ↑Wikipedia. (2026). Kilo Code. https://en.wikipedia.org/wiki/Kilo_Code
  2. ↑Kilo Code. (2025). We raised $8 Million to bring Kilo Speed to Agentic Engineering. https://blog.kilo.ai/p/kilo-raised-8-million-seed-round
  3. ↑Kilo-Org. (2026). kilocode source repository. https://github.com/kilo-org/kilocode
  4. ↑Agentic Index. (2026). Kilo Code vendor assessment (Anaconda acquisition, July 2026). https://agenticindex.io/vendors/kilo-code
  5. ↑Malaysia Digital Economy Corporation. (2026). Official website. https://www.mdec.my/
Back to all articles