Kilo Code

What is Kilo Code?

Kilo Code is an open-source AI programming and agent orchestration platform that can be used in VS Code, JetBrains, the terminal, web-based cloud proxies, and mobile devices. It integrates code generation, planning, debugging, testing, code review, application building, and deployment within a single account and session framework.

Kilo is not tied to any single model provider. Users can access a large number of models through the Kilo Gateway; they can also connect to their own accounts such as those of Anthropic, OpenAI, Google, and AWS Bedrock, or use Ollama and LM Studio to run local models.

New version 1.0 architecture

Current versions of Kilo Code 1.0 and later use the Kilo CLI as the core technology to unify the experience in the IDE and the terminal. The VS Code extensions have been rebuilt on this same foundation, allowing the CLI, extensions, and other platforms to share models, proxies, and session capabilities.

On July 31, 2026, support for the older versions of VS Code and JetBrains will come to an end; the old repositories will be archived in August.

When installing and troubleshooting, make sure to use the current kilocode repository and the latest documentation; do not continue to rely on the old configuration guides.

VS Code extensions

The VS Code version provides features such as proxy settings, context references, file editing, terminal execution, autocompletion, diff comparison, and task management in the sidebar. It can also run in editors that support VS Code extensions, such as Cursor and Windsurf.

Agents can modify the entire project and execute commands. For important repositories, Git branches should be used, automatic approvals should be restricted, and the differences should be checked before applying any changes.

JetBrains plugins

Kilo is compatible with JetBrains products such as IntelliJ IDEA, PyCharm, and WebStorm, offering features like chat, code manipulation, proxy mode, and model selection. The new version of this plugin is gradually being integrated with the CLI infrastructure, making it suitable for Java, Kotlin, and Python teams that do not use VS Code.

The auto-completion, code manipulation features, and keyboard shortcuts vary among different IDEs; therefore, it is necessary to test the target version in the remote development environment before deploying it to a team.

Kilo CLI

Kilo CLI offers a terminal TUI that enables the planning and execution of code tasks, switching between models and agents, managing MCP, running headless tasks, connecting to GitHub, checking PRs, synchronizing remote sessions, and viewing token costs. The official installation package also supports global installation via npm.

CLI 1.0 requires a modern operating environment; if an older CPU lacks AVX, it is necessary to download the baseline version. The automatic mode disables permission prompts, and it should only be used in isolated and trusted environments.

Code proxy

The Code agent is used to implement functions, modify multiple files, execute commands, and carry out tests; it represents the main approach in everyday development. It enables continuous iteration by combining the current repository with user instructions.

The objectives should include acceptance criteria and the areas that are not to be modified. Providing the model with overly vague requirements at once increases the likelihood of errors and raises the costs associated with reasoning.

Plan and Architect agents

Plan or Architect is suitable for reading through a project, analyzing its impacts, formulating implementation plans, and breaking down tasks, before handing over the clear steps to the execution agents. For large-scale migrations and new modules, planning first usually makes it easier to control risks.

Planning is merely the result of reasoning; it is necessary to examine the actual code, dependencies, and business constraints, and one should not treat a well-designed plan as proof of technical viability.

Ask and Debug proxies

Ask is used to understand code and answer technical questions, while Debug helps identify the root causes by examining logs, the terminal output, and execution results. Users can first conduct a read-only investigation, and then grant the agent permission to modify the code.

During debugging, the original errors and the simplest way to reproduce them should be retained. When the model cannot make a decision based on evidence, it should be instructed to run verification commands rather than trying various modifications repeatedly.

Orchestrator and multiple agents

Kilo includes built-in proxy orchestration capabilities that allow complex tasks to be broken down into multiple individual tasks; the planning, coding, debugging, and review proxies can then be invoked in parallel or sequentially. The new version integrates some of these orchestration functions into all proxies, while the old Orchestrator approach can still be used for backward compatibility with existing workflows.

Parallel agents increase the number of calls and lead to merge conflicts. Each branch should have clear responsibilities, isolated work areas, and defined steps for final integration.

Custom proxies and Skills

Users can create custom proxies, commands, and Skills, thereby encapsulating framework specifications, testing steps, deployment processes, and domain-specific knowledge into reusable capabilities. The CLI can also manage these configurations through proxy and plugin commands.

Shared configurations should not contain keys or customer data. Skills should define permissions, tools, and inputs to prevent any task from automatically triggering high-risk operations.

Auto-completion

Kilo offers in-line code completion within the IDE; models such as Mistral and Codestral are currently available for use. Users can employ their own Codestral keys, or they can run the functionality through the Kilo account settings.

Autocomplete may continuously send requests in the background. If not needed, this feature can be turned off.

The team should also verify the data processing terms of the selected model.

Contextual search and Memory Bank

Automatic context search helps agents locate relevant files, functions, and symbols within the code base, while the Memory Bank is used to store project conventions, architecture details, and long-term information, thereby reducing the need for repeated interpretation.

Memories may become outdated or conflict with code; it is necessary to indicate their source and update date, and to clean them up regularly. Sensitive information should not be treated as ordinary shared memories.

Checkpoints – points for verification

Kilo can save the working state before and after task execution, allowing users to compare different versions and revert any unsatisfactory changes. It is suitable for exploratory refactoring and multiple rounds of iterative modifications.

Checkpoints cannot replace remote Git repositories, commits, or backups. Before restoring, it is necessary to ensure that no unsaved manual changes made by users are overwritten.

MCP and external tools

The platform supports the Model Context Protocol; proxies can connect to documents, databases, browsers, Issues, monitoring systems, and custom services. The CLI provides MCP management commands, and projects can also share approved server configurations with their teams.

The MCP tool is capable of reading from or modifying external systems. An allowlist, environment variables, minimum permissions, and approval processes should be established; in addition, it is prohibited to store administrator tokens in the repository.

Used by browsers

Kilo offers browser automation capabilities for web page debugging, screenshotting, front-end validation, and form handling. Cloud Agent and App Builder can also run actual applications and check their performance.

Websites may contain malicious prompts; before clicking, logging in, or submitting, it is necessary to verify the target domain name and the relevant permissions. Payments, deletions, and public publications should not be approved automatically.

App Builder

App Builder enables users to create full-stack applications from natural language descriptions; they can further modify the user interface, the backend, and the data logic, and preview the resulting application in a cloud environment. It is suitable for creating prototypes, internal tools, and marketing pages.

When creating an application, it is still necessary to check authentication procedures, database permissions, keys, potential vulnerability issues, compatibility with mobile devices, and accessibility; the application cannot be launched directly just because the preview works properly.

Cloud Agent

Cloud Agent executes code tasks within isolated containers; users can keep it running even after turning off their local computers, and they can access the sessions from web browsers, mobile devices, or other platforms. It is suitable for long-running tasks, parallel development, and remote task delegation.

The cost of cloud proxies includes both model inference and container computing. Once a task is completed, the container should be stopped, and access to repositories, cloud credentials, and the network should be restricted.

Code review

Kilo Code Reviewer can analyze pull requests, identify issues, provide suggestions, and make automatic corrections. The CLI also allows for pulling PR branches and running the review agent locally.

AI-based review cannot replace human approval, static scanning, and CI. It may fail to detect business risks or generate stylistic issues; therefore, project rules and severity levels should be established.

Deploy deployment

Kilo Deploy enables the rapid deployment of projects such as Next.js and static websites, offering an integrated process that spans from code generation to preview and deployment. Teams can run tests and security checks before deploying.

Production deployment involves domain names, environment variables, billing, and rollback procedures. It is necessary to identify the target environment and maintain manual controls to prevent proxies from deploying test configurations to production.

Sessions cross-device sessions

Sessions can be continued across the IDE, CLI, cloud, and mobile devices; users can share, rename, export, import, and view timelines. This allows tasks to be initiated on a smartphone while the code is reviewed on a desktop.

Sharing sessions may expose hints, code snippets, and tool results. It is necessary to check the content and determine how to revert it before sharing externally.

Supported models

Kilo Gateway currently supports over 60 providers and more than 500 models, including Claude, GPT, Gemini, Grok, DeepSeek, Kimi, MiniMax, as well as open-source models. Users can also use their own provider keys directly.

The model directories and Auto-routing are updated dynamically; it is not possible to define the underlying models based on old documents. Production tasks should record the actual model IDs and versions.

Auto Model automatic routing

Auto Model allows selection of models based on categories such as Frontier, Efficient, Free, etc., as well as the current proxy mode; this enables the use of more powerful models for complex tasks while helping to control costs for simpler tasks. The routing mapping is updated by the server.

Automatic selection is convenient, but it also reduces reproducibility. For rigorous audit tasks, the model should be fixed, and tokens, parameters, and output records should be saved.

Free usage method

The personal platform itself is free; users can set the proxy interactions, completion functions, and backend summaries to use free models, thereby avoiding the need to purchase hosting credits. It is also possible to use Ollama, LM Studio, or custom APIs that come with free credits.

The anonymous free version allows 200 requests per hour, limited by IP address; certain free endpoints log data for security purposes or to improve the service, and they are not suitable for use with sensitive code.

Price and version comparison

Kilo Gateway is billed on a pay-as-you-go basis.

Gateway calculates the tokens for input, output, cache writes, and cache hits based on the unit prices published by upstream providers; it does not add any extra charge to the model’s price. A payment processing fee of 5% is applied to top-ups, and the purchased quota remains valid indefinitely.

When the account balance is insufficient, payment requests will be halted. Organizations can set a shared balance, a daily limit per user, automatic top-ups, and alerts for low balances.

Price of Kilo Pass

Kilo Pass converts the monthly fee into an equivalent amount of credit, and awards additional credits based on the duration of the subscription. The Starter plan costs 19 dollars per month, with a maximum credit amount of around 26.60 dollars.

The Pro plan costs 49 dollars per month, with a maximum of around 68.60 dollars; the Expert plan costs 199 dollars per month, with a maximum of around 278.60 dollars.

The bonus for the first month of a monthly subscription can be as high as 50%, and it rises to 40% with continuous subscriptions; for an annual subscription, the bonus is 50% per month.

The reward amount expires at the end of the month, while the payment quota remains valid; the actual rewards are indicated on the account page.

Team and enterprise prices

The individual plan costs $0. Teams currently costs $15 per user per month; it offers team management, shared proxies, centralized billing, analytics, adoption scoring, shared BYOK, privacy controls, and priority support, along with a 14-day trial period.

Enterprise offers quotes on demand, including additional features such as SSO, OIDC, SCIM, audit logs, restrictions related to models and providers, a private gateway, SLAs, and dedicated support. The seat fee does not include any model quotas.

Cloud Agent pricing

Cloud computing and model inference are billed separately. Currently, Docker and Small Cloud Agent cost 0.60 dollars per hour, Standard costs 1.20 dollars per hour, Code Review costs 0.33 dollars per hour, and Gas Town also costs 1.20 dollars per hour; the charges are usually calculated on a per-second basis.

Computing costs are incurred during the container’s activation, and calling the model also consumes credits. Long-running tasks should have stop conditions and budget limits set.

AI Gateway API

Kilo Gateway offers OpenAI-compatible interfaces, streaming responses, tool calls, a list of models, cost tracking, and BYOK. Developers can use the OpenAI SDK, Vercel AI SDK, or other compatible clients to access the same model directory.

The validity period of organizational tokens and the related policies are controlled by the account. API keys should be stored in a secret management system; they must not be included in the code or in the frontend components.

GitHub and open-source migration

The current source code for Kilo Code is located in the kilocode repository on Kilo-Org; it includes code related to CLI, VS Code, and JetBrains. It is licensed under the MIT license and builds upon the OpenCode technology branch. Users are allowed to view, modify, and use it for commercial purposes, provided that they retain the license statement.

The old kilocode-legacy repository uses Apache 2.0 and has been archived; it is retained only for historical reference. The current repository should be used when determining the open-source license and when submitting issues.

Privacy and Corporate Governance

When using managed Gateways and Cloud Agents, the code, prompts, tool outputs, and usage information are sent to the respective services. BYOK keys are used for encryption, and organizations can set restrictions on models, providers, data collection, and daily expenses.

Enterprise’s private Gateway, EU-based inference capabilities, and built-in computing solutions are suitable for more stringent use cases; however, it is still necessary to conduct evaluations regarding contracts, data location, and access control.

Kilo Code usage guide

Complete a basic task.

1. Define the code tasks to be completed in Kilo Code, the scope of the repository, and the acceptance criteria.
2. Connect to or import the test project, and first back up the current branch;
3. First, generate the plan using the new version 1.0 architecture, and then confirm the files that need to be modified.
4. Use VS Code extensions to make small changes;
5. Run tests, static checks, and builds; unverified code is not accepted directly.
6. Manually check permissions, keys, dependencies, and handle exceptions before merging;

Create reusable professional workflows

1. Select a low-risk, real-world project as a template;
2. Fix the order of using the new version 1.0 architecture, VS Code extensions, and JetBrains plugins;
3. Record the environment, model, prompts, and failure conditions;
4. Set up manual approval for write, deploy, and delete actions;
5. Compare speed, cost, test pass rate, and the amount of rework;
6. Expand to a team or production environment only after stability has been verified;

Which users is it suitable for?

• Developers who wish to use AI agents consistently in VS Code, JetBrains, and the terminal;
• Teams that need to choose between multiple models, BYOK, or local models;
• Engineering organizations that require parallel Cloud Agents, code reviews, and application building;
• Advanced users who place emphasis on open source, cross-device sessions, and proxy orchestration;
• Enterprises that require analysis, SSO, auditing, and model governance.

Product advantages

• The current core repository is open-source under MIT;
• Covers IDE, CLI, the cloud, mobile devices, and Slack;
• Access to over 500 models is available, with support for both local use and BYOK;
• The costs related to platforms, inference, and cloud computing are clearly separated and displayed;
• It features multiple agents, review capabilities, deployment options, and an App Builder;
• An organization can set model, data, and spending strategies.

Restrictions and Precautions

• Product features and versions are updated rapidly, and old extensions, configurations, and repositories are no longer maintained.
• The tutorial must indicate that it is suitable for version 1.0 and later;
• Agents can execute commands, browse web pages, and deploy applications; costs related to cloud services and models may also apply.
• Minimum permissions, budget controls, Git review, testing, and manual release approvals should be employed;

Frequently Asked Questions

Is Kilo Code free?

The personal platform is free to use; users can opt for free models, local models, or BYOK. Paid models, Kilo Pass, and cloud computing services are charged according to their respective pricing rules.

Which editors does Kilo Code support?

It supports VS Code, JetBrains, and some VS Code-compatible editors; it also provides a standalone CLI.

How much is Teams?

The current rate is $15 per user per month, and this does not include the quota for model inference; contact sales for an Enterprise quote.

How much is Kilo Pass?

Starter costs $19 per month, Pro costs $49 per month, and Expert costs $199 per month; bonus amounts are awarded in accordance with the rules.

Does Kilo Code support local models?

It supports Ollama and LM Studio, and it is also possible to connect to custom cloud model keys.

Is Kilo Code open source?

It is open source; the current kilocode repository uses the MIT license, while the old Apache 2.0 repository has been archived.