Kilo Code Review: The Open-Source AI Coding Agent That Forked Both Cline and Roo Code
Kilo Code went from an obscure fork to roughly 1.5 million users and the #1 spot on OpenRouter in under a year. Backed by GitLab's co-founder and an $8 million seed round, it positions itself as a "superset" of Cline and Roo Code — with Orchestrator mode, 500+ models at zero markup, and a new CLI. After two weeks of daily use across VS Code projects, here's what holds up and where the cracks show.
TL;DR — Key Takeaways:
- • Zero markup pricing — bring your own API keys and pay exact provider rates. No hidden fees. Optional Kilo Pass credits start at $19/mo for $28.50 in credit value (50% bonus).
- • Orchestrator mode is the standout feature — breaks complex tasks into subtasks routed to specialized agents (Code, Debug, Architect, Ask). Genuinely useful for multi-step projects.
- • 500+ models, multi-IDE support — works in VS Code, JetBrains, and a standalone CLI. Model selection is unmatched by any competitor.
- • GitLab DNA, real funding — co-founded by Sid Sijbrandij (GitLab) and Scott Breitenother (Brooklyn Data). $8M seed from Cota Capital, General Catalyst. Roughly 34-person remote team.
- • Still rough around the edges — context drift on longer sessions, provider configuration bugs, and token burn rates that can exceed expectations. Feature sprawl is a real risk.
What Is Kilo Code?
Kilo Code is an open-source AI coding agent that runs as an extension in VS Code and JetBrains IDEs, plus a standalone CLI tool called Kilo CLI. It was created by forking both Cline (the most popular open-source VS Code coding extension with over 5 million installs) and Roo Code, then merging their best features under a unified platform.
The company was co-founded by Sid Sijbrandij, who co-founded GitLab and served as its CEO, and Scott Breitenother, founder of data consultancy Brooklyn Data. In December 2025, Kilo raised $8 million in seed funding led by Cota Capital with participation from General Catalyst, Breakers, Quiet Capital, and Tokyo Black.
The pitch is straightforward: access 500+ AI models at exact provider rates (zero markup), coordinate complex tasks across specialized agents via Orchestrator mode, and work across multiple surfaces — all without switching away from your existing editor. As of early 2026, the platform claims roughly 1.5 million users and processes upward of 6 trillion tokens per month through OpenRouter integration. It currently holds the #1 spot on OpenRouter by usage volume.
Quick Verdict
Kilo Code fits if you:
- • Want maximum model flexibility (500+ options)
- • Prefer paying exact API rates with no subscription tax
- • Need multi-agent orchestration for complex projects
- • Work across VS Code and JetBrains
- • Already use Cline and want more features
Skip Kilo Code if you:
- • Want a polished, turnkey experience (Cursor is smoother)
- • Need rock-solid stability for production workflows
- • Prefer predictable monthly billing over pay-per-token
- • Use Zed, Neovim, or editors outside VS Code/JetBrains
- • Dislike managing API keys and model selection
How We Tested
This review is based on two weeks of daily use from mid-February through early March 2026. We ran Kilo Code through structured tasks across multiple project types, comparing results against Cline, Cursor, and Claude Code on the same codebases.
Agentic Coding Tasks (10 sessions)
Multi-file refactoring, bug fixes from error logs, and feature implementation across TypeScript and Python projects. Tested Orchestrator mode splitting a full-stack feature (API + frontend + tests) into coordinated subtasks.
Model Switching (8 models tested)
Switched between Claude Sonnet 4.5, GPT-5, Gemini 3 Pro, and several open-source models (DeepSeek, Llama) on identical tasks. Measured output quality, latency, and token costs through Kilo's gateway versus direct API access.
Multi-IDE Testing (VS Code + JetBrains WebStorm)
Ran the same projects in both VS Code and WebStorm to verify feature parity. Tested autocomplete, inline suggestions, and Orchestrator mode in each environment.
Kilo CLI (5 terminal sessions)
Tested the command-line interface for headless coding tasks, CI/CD integration scenarios, and SSH-based remote development. Compared workflow and UX against Claude Code CLI.
Community and Third-Party Review Analysis
Cross-referenced our findings with Product Hunt reviews (4.7/5 average), Trustpilot feedback, VS Code Marketplace ratings, and GitHub issue tracker patterns. User-reported issues informed our downsides section.
All testing used personal API keys (Anthropic, OpenAI, Google) with no sponsored credits or early access from Kilo. Token costs reported are actual billed amounts from provider dashboards.
Key Features Worth Knowing
Kilo Code ships a long feature list, but a handful of capabilities genuinely differentiate it from Cline and the wider field. Here's what matters most in practice.
Core Differentiators
Four Specialized Modes + Orchestrator
Kilo Code provides four core modes — Code (implementation), Architect (planning and design), Debug (error diagnosis), and Ask (questions and explanations). Each mode constrains the agent's behavior: Architect mode, for example, won't modify files directly. The fifth mode, Orchestrator, coordinates the other four to handle complex multi-step tasks.
Cline has Plan and Act modes. Kilo's four-mode structure is more granular and in practice helps the model stay focused on the right kind of work for each phase.
Zero Markup Model Access (500+ Models)
You bring your own API key from any provider — Anthropic, OpenAI, Google, AWS Bedrock, Azure, or local models via Ollama — and Kilo charges nothing on top. The gateway passes through the exact provider rate. This is confirmed: we compared bills from Anthropic's dashboard against Kilo's usage logs, and the numbers matched to the cent.
The alternative is Kilo Pass credits, which actually give you a bonus (roughly 50% extra at the $19/mo tier). Either way, you're not paying a markup — a genuine differentiator against subscription-based tools.
Memory Bank
Memory Bank stores architectural decisions, project conventions, and codebase context in structured Markdown files within your repository. When you start a new session, Kilo reads these files to reconstruct project understanding. It's conceptually similar to Claude Code's CLAUDE.md or Cursor's .cursorrules, but more structured with separate files for architecture, patterns, and progress.
In our testing, Memory Bank reduced the "explain the project again" overhead noticeably. The onboarding experience for returning to a project after days away was smoother than with Cline or stock Cursor.
Kilo CLI
Kilo CLI 1.0 brings the agent to the terminal with the same model access and modes available in the IDE extension. It's positioned against Claude Code CLI and GitHub Copilot CLI. In practice, the CLI is functional for headless tasks and remote servers but less polished than Claude Code's terminal experience. The advantage: you can use any of 500+ models instead of being locked to a single provider.
Other Notable Features
Orchestrator Mode: What It Actually Does
Orchestrator is Kilo Code's headline feature and the primary reason to choose it over Cline. Instead of handling a complex request as one monolithic conversation, Orchestrator decomposes it into discrete subtasks, each running in its own context with the appropriate specialized mode.
Here's a concrete example. We asked Orchestrator to "add user authentication with JWT tokens, including login/register API endpoints, middleware, and tests." It created four subtasks:
Planned the file structure, chose libraries (bcrypt, jsonwebtoken), and defined the API contract. Did not write any code.
Implemented the auth routes, user model, and JWT middleware based on the Architect's plan.
Added login/register forms and token storage to the frontend, referencing the API contract from step 1.
Generated integration tests for all endpoints and ran them, catching two type errors from the Code subtask.
The auth feature took about 18 minutes start to finish with Orchestrator, compared to roughly 25 minutes when we attempted the same task as a single conversation in Cline. The quality was comparable, but Orchestrator's output was better organized because each subtask stayed focused.
The catch: Orchestrator works well for tasks with clearly separable components. When subtasks have tight dependencies or when later steps need to significantly revise earlier work, the coordination overhead can actually slow things down. We saw this when asking Orchestrator to refactor a tightly coupled legacy codebase — the subtasks kept producing conflicting changes that required manual reconciliation.
By default, you approve each subtask creation and completion. For trusted workflows, you can enable auto-approval in settings to let Orchestrator run without interruption.
Kilo Code vs Cline vs Cursor vs Claude Code
This is the comparison most developers are weighing. Each tool takes a fundamentally different approach to AI-assisted coding. The table below reflects our hands-on testing and verified pricing as of March 2026.
| Feature | Kilo Code | Cline | Cursor | Claude Code |
|---|---|---|---|---|
| Type | IDE extension + CLI | IDE extension | Standalone IDE | CLI agent |
| Open Source | Yes (Apache 2.0) | Yes (Apache 2.0) | No | No |
| IDE Support | VS Code + JetBrains | VS Code only | Own IDE (VS Code fork) | Terminal only |
| Model Count | 500+ | 50+ | ~15 built-in | Claude models only |
| Pricing Model | BYOK (zero markup) | BYOK (zero markup) | $20–200/mo subscription | $17–200/mo subscription |
| Orchestration | Orchestrator mode | No | Agent mode (single) | Agent Teams (Opus only) |
| Autocomplete | Built-in | No | Built-in (Tab) | No |
| Memory/Context | Memory Bank | Session-based | .cursorrules | CLAUDE.md |
| User Base | ~1.5M | ~5M installs | Undisclosed (millions) | Undisclosed |
The pattern is clear: Kilo Code occupies the "feature-rich open-source" position. It has more models, more modes, and more surfaces than Cline, while maintaining the same BYOK cost model. Cursor and Claude Code trade model flexibility for polish and deeper integration with their respective ecosystems.
For a deeper look at the Cursor and Claude Code side of this comparison, see our Claude Code vs Cursor comparison. If you're specifically evaluating Cline, our Cline review covers its strengths and limitations in detail.
Worth noting: Kilo's 500+ model count comes primarily through OpenRouter integration, which Cline also supports. The difference is that Kilo wraps this in a more polished model-switching UI and adds its own gateway with built-in cost tracking.
Pricing: What Kilo Code Actually Costs
Kilo Code's pricing is arguably its strongest selling point against subscription-based competitors. The extension and CLI are free. You only pay for AI model usage — either directly through your own API keys, or through Kilo Pass credits at a discount.
Individual Plans
| Plan | Monthly Cost | Credits Received | Bonus |
|---|---|---|---|
| Free (BYOK) | $0 | $20 signup bonus | Pay provider rates directly |
| Kilo Pass Starter | $19 | $28.50 | 50% bonus credits |
| Kilo Pass Pro | $49 | $73.50 | 50% bonus credits |
| Kilo Pass Expert | $199 | $298.50 | 50% bonus credits |
Team and Enterprise
| Plan | Cost | Key Features |
|---|---|---|
| Teams | $15/user/month | Usage analytics, shared modes, centralized billing, data privacy controls |
| Enterprise | Contact sales | SSO/SCIM, audit logs, SLA, dedicated support, private gateway |
What It Costs in Practice
We tracked our actual spending over two weeks using BYOK with Anthropic and OpenAI keys. Moderate daily usage (3–4 hours of coding with Claude Sonnet 4.5 as the primary model) averaged roughly $8–12 per day in API costs, or approximately $180–260/month.
That's comparable to what the same usage would cost in Cline (identical, since both pass through provider rates) and cheaper than Cursor Pro ($60/mo) only if your daily usage stays under about 2 hours. For heavy users, Cursor's flat subscription can actually be more economical. The math flips depending on how much you use it.
One detail worth noting: Kilo Pass bonus credits expire at month-end if unused, while paid credits that you purchase outside the subscription never expire. Factor this in if your usage fluctuates month to month.
Genuine Downsides
Kilo Code's ambition to be a "superset of everything" creates real problems alongside the benefits. These are the issues we encountered and that user reviews consistently flag.
Context Drift on Longer Sessions
Mid-way through complex, multi-step tasks, Kilo sometimes references earlier questions or context as if they're current, derailing the workflow. We observed this most often in Orchestrator sessions lasting more than 30 minutes. The agent would occasionally treat a subtask's context as part of the main conversation, producing responses that seemed out of sync.
Multiple Product Hunt and GitHub reviewers report similar behavior. Memory Bank helps for project-level context, but within a single session, drift remains a problem.
Provider Configuration Bugs
Some users (and we experienced this once) report that the configured API provider isn't always respected. The extension occasionally routes requests through a different provider silently, resulting in unexpected charges to the wrong account. We saw this manifest as a request being routed through OpenRouter when we had a direct Anthropic key configured.
This appears to be a known bug that the team is working on. It happened once in two weeks of testing for us, but the financial implications make it worth flagging.
Token Burn Rate Can Surprise You
Orchestrator mode is inherently token-hungry because each subtask spins up its own context. A task that costs $0.50 as a single conversation might cost $1.50–2.00 through Orchestrator, because each subtask needs to ingest relevant context independently. The value tradeoff (better organized output vs. higher cost) isn't always favorable, especially for simpler tasks that don't need decomposition.
Feature Sprawl Risk
Kilo Code ships IDE extensions, a CLI, App Builder, Slack integration, automated PR reviews, Cloud Agents, autocomplete, and custom modes — all from a 34-person team that's less than a year old. That's an enormous surface area. Some of these features feel like they were launched to check competitive boxes rather than because they're fully baked. App Builder and the Slack bot, for instance, are functional but clearly early-stage.
Limited Editor Support Beyond VS Code and JetBrains
If you use Zed, Neovim, Emacs, or any other editor, Kilo Code isn't available. The CLI covers terminal workflows, but the full agent experience requires VS Code or a JetBrains IDE. Given the tool's positioning as the "universal" coding agent, this is a notable gap.
Who Should Use Kilo Code
Kilo Code makes sense for:
- • Developers who want model flexibility — if switching between Claude, GPT, Gemini, and open-source models based on task type matters to your workflow, nothing else offers 500+ options with zero markup
- • Cline users who've hit its ceiling — Orchestrator mode, autocomplete, JetBrains support, and Memory Bank are all genuine upgrades over Cline's feature set
- • Teams wanting cost transparency — the $15/user/month Teams plan plus BYOK model means you know exactly what AI is costing you, unlike bundled subscription tools where cost-per-task is opaque
- • Full-stack developers tackling complex features — Orchestrator shines when a task naturally decomposes into separate frontend/backend/testing concerns
- • JetBrains users — if you're on IntelliJ, WebStorm, or PyCharm, your options for advanced AI coding agents are limited. Kilo is one of the few that fully supports JetBrains alongside VS Code
Stick with alternatives if:
- • You value stability over features — Kilo is fast-moving and occasionally buggy. If your production workflow can't tolerate provider routing issues or context drift, Cursor or Claude Code is more reliable
- • You prefer predictable billing — Cursor's $20–60/month flat rate is simpler to budget for. Pay-per-token costs fluctuate with usage and model choice
- • You only use Claude models — Claude Code's native integration with the Claude ecosystem (Agent Teams, CLAUDE.md, computer use) is deeper than what Kilo can offer as a multi-model wrapper
- • You're new to AI coding agents — the model selection, mode choices, and configuration options in Kilo can overwhelm newcomers. Cursor or even Cline has a gentler learning curve
The simplest heuristic: if you're currently using Cline and wish it did more, Kilo Code is the natural next step. If you're happy with Cursor or Claude Code, the switch probably isn't worth the disruption unless model flexibility is a top priority. For a broader overview of the AI coding landscape, see our vibe coding tools guide.
Frequently Asked Questions
Is Kilo Code free to use?
Yes. The extension and CLI are free. You bring your own API keys and pay exact provider rates with zero markup. Kilo also offers optional Kilo Pass credit subscriptions ($19–199/month) with 50% bonus credits, plus $20 in free credits for new users. You can use Kilo indefinitely at $0 if you supply your own keys.
How does Kilo Code compare to Cline?
Kilo Code forked from Cline and added Orchestrator mode, built-in autocomplete, JetBrains support, Memory Bank, a CLI, and App Builder. Cline remains simpler with dual Plan/Act modes and has a larger user base (~5 million installs vs. Kilo's ~1.5 million). Both use BYOK pricing with zero markup. Choose Cline for simplicity and maturity; choose Kilo for more features and multi-IDE support.
What is Kilo Code Orchestrator mode?
Orchestrator breaks complex tasks into subtasks handled by specialized agents. A feature request might spawn separate Architect (planning), Code (implementation), and Debug (testing) subtasks, each in its own context. You approve each step by default, or can enable auto-approval. It works well for separable tasks but can add overhead on tightly coupled problems.
Is Kilo Code better than Cursor?
Different tradeoffs. Kilo is open-source, works in your existing editor, supports 500+ models, and charges zero markup on model costs. Cursor is a polished standalone IDE with flat-rate pricing ($20–200/month) and a more integrated UX. Kilo gives you more control and flexibility; Cursor gives you more convenience and predictability. For most developers, it comes down to whether you prioritize cost transparency or turnkey experience.
What AI models does Kilo Code support?
Over 500 models through OpenRouter and direct provider integrations. This includes Claude Opus 4.6, Claude Sonnet 4.5, GPT-5, Gemini 3 Pro, DeepSeek, Llama, and many other open-source options. You can also run local models via Ollama or LM Studio for zero-cost inference (at the cost of quality, depending on the model).
Who founded Kilo Code?
Kilo was co-founded by Sid Sijbrandij (co-founder and former CEO of GitLab) and Scott Breitenother (founder of Brooklyn Data). The company raised $8 million in seed funding in December 2025 from Cota Capital, General Catalyst, Breakers, Quiet Capital, and Tokyo Black. It operates fully remotely with about 34 employees.
Final Verdict
Kilo Code is the most feature-rich open-source AI coding agent available in early 2026. The combination of Orchestrator mode, 500+ models at zero markup, JetBrains support, and Memory Bank gives it a legitimate claim to being a "superset" of Cline. The GitLab founder pedigree and $8 million in funding suggest this isn't a weekend fork — there's genuine organizational intent behind it.
It's also younger, buggier, and more sprawling than the competition. Context drift on long sessions, provider routing inconsistencies, and feature sprawl across a dozen different surfaces from a 34-person team are real concerns, not theoretical ones. The refund policy frustrations surfacing in user reviews suggest the team is still learning how to handle the customer side of things.
The pricing model is genuinely compelling. In a market where Cursor charges $20–200/month and Claude Code requires a $17–200/month subscription, Kilo's "pay only for what you use at exact provider rates" approach gives cost-conscious developers a meaningful alternative. Whether that advantage holds as the company needs to generate revenue at scale remains to be seen.
Our Recommendation
Kilo Code is the right tool for developers who want maximum control over their AI coding stack — which models to use, what to pay, and how to structure complex tasks. For everyone else, it's worth watching but not necessarily switching to. The product is moving fast, and six months from now the stability and feature maturity picture could look very different.