Roo Code has been shut down. The project's README says the extension was shut down on 15 May, and the GitHub repository sits in the archive. So the question "how do Kilo Code and Roo Code differ from Cline" comes down to choosing between two living extensions, with the third left to a footnote. Cline and Kilo Code accept any OpenAI-compatible endpoint, and codexkey needs four values: the address https://api.codexkey.online/v1, a key that starts with sk_cdx_, a model ID and a context size.
What follows is the comparison, setup for each extension, model choice, token spending and the typical errors. Project status is given as of 10 November 2026, from the READMEs and docs of Cline, Roo Code and Kilo Code opened that day. The extensions were not run on a clean machine, so wherever a settings screen comes from the docs and not from a run, the text says so.
Cline, Kilo Code and Roo Code side by side
| Cline | Kilo Code | Roo Code | |
|---|---|---|---|
| License | Apache 2.0 | MIT | Apache 2.0 |
| Where it runs | VS Code, a JetBrains plugin, a CLI, a desktop app | VS Code, a JetBrains plugin, a CLI, a cloud agent in the browser | VS Code only |
| Modes | Plan and Act | Code, Plan, Ask, Debug and custom ones | Code, Architect, Ask, Debug and custom ones |
| Origin | The source project: Roo Code branched off it | Began as a fork of Roo Code, the current extension is built on the Kilo CLI core | Came out of Cline |
| Status on 10 Nov 2026 | Active | Active, Anaconda announced it is buying the company | Shut down 15 May 2026, repository archived |
| Custom endpoint | OpenAI Compatible provider | Custom provider dialog | Used to be OpenAI Compatible |
Licenses, modes and platforms come from each project's README.
The projects treat money differently. Both living extensions are open, and you pay for tokens either through the project's account or straight to the model provider. Cline's docs describe three ways in: an account sign-in, a subscription and your own key. Kilo promises provider pricing with no markup and more than 500 models through its account, while your own key works with more than 30 providers. The last path is the one that matters here, because you enter the codexkey key yourself.
Where Kilo came from is visible in its April 2025 blog post: the team forked Roo Code, which was itself a branch of Cline. Kilo has since moved far. According to its docs, today's extension is built on the Kilo CLI, and the README says that CLI is a fork of OpenCode. By its code you can no longer call it a lightly modified Cline.
Kilo's ownership changed too. On 15 July 2026 Anaconda announced that it is acquiring Kilo Code, and the documentation pages now carry a banner saying the deal is done. The announcement promises that Kilo stays available to developers and teams. It says nothing about the license or pricing.
What happened to Roo Code
The Disclaimer section in Roo Code's README says the extension was shut down on 15 May and points anyone seeking a replacement to Cline and ZooCode. On GitHub, RooCodeInc/Roo-Code is marked archived, with its last change dated the same 15 May 2026. Roo's documentation site is still up, and it carries the same shutdown banner.
ZooCode's own README calls it a fork started by the Roo Code community. It states the Apache 2.0 license, VS Code as the main platform, the familiar Code, Architect, Ask, Debug and custom modes, and a migration guide from Roo. ZooCode was not installed, so nothing can be vouched for about its stability.
Already have Roo Code in the editor? Its provider fields match Cline's: Roo's docs list API Provider, Base URL, API Key, Model and a Model Configuration block with a Context Window field. Fill them in with the Cline steps below. The project will get no fixes, though, so for a fresh install take Cline or Kilo Code.
Two connection formats, any extension
All three extensions share one scheme, which the extension guide in codexkey's docs also describes: the provider type is either OpenAI Compatible or Anthropic. The choice decides whether the address ends in /v1.
| Provider type | Base URL | Key |
|---|---|---|
| OpenAI Compatible | https://api.codexkey.online/v1 | your sk_cdx_... key |
| Anthropic | https://api.codexkey.online | your sk_cdx_... key |
OpenAI Compatible serves Claude, GPT and the other catalog models under one key. The Anthropic mode keeps the Claude format and is needed less often: in Cline it is the Anthropic provider with the Use custom base URL option, in Kilo it is the Anthropic Messages item in the Provider API list. How Kilo treats the address in that mode was not checked, so the steps below use OpenAI Compatible.
Step 1. Get a codexkey key
Sign up, top up your balance and create a key in the API Keys section of the dashboard. The minimum top-up is $1, and the top-up dialog shows the exact limits for the method you choose. Rubles through SBP (the Russian Faster Payments System), Russian cards and crypto (USDT, BTC, ETH) are all accepted. The key looks like sk_cdx_..., and one key opens both Claude and GPT.
Do you need it at all? If Kilo already runs through its own account with credits, stay with that. A codexkey key makes sense when you want Claude models without an Anthropic account, one balance shared by Cline, Claude Code and Cursor, or a payment route that works from where you live.
Step 2. Set up Cline
In the Cline panel click the gear and choose OpenAI Compatible in the API Provider field. Fill in the other fields as the guide has them:
API Provider: OpenAI Compatible
Base URL: https://api.codexkey.online/v1
API Key: sk_cdx_YOUR_KEY
Model ID: claude-sonnet-5
Context Window Size: 1000000
In Cline's docs the context size and the output limit are gathered in a Model Configuration block, next to switches for image support and per-token prices; none of that is needed to get going. The [1m] suffix does not belong in the model ID: it was made for Claude Code, and every other program takes the name from the catalog as it stands.
If you want Claude only and the Anthropic format, pick the Anthropic provider, turn on Use custom base URL and enter the address without /v1. A longer walkthrough with a curl check is in Setting up Cline in VS Code with a custom base URL.
Step 3. Set up Kilo Code
Kilo's screen is different, and the description below rests on Kilo's documentation, not on a run. The gear in the extension panel opens settings; on the Providers tab scroll the list to the end and click Custom provider. The dialog asks for these fields:
- Identifier (Provider ID): any unique string, for example
codexkey. - Name (Display name): the label you will see in the interface.
- API type (Provider API): OpenAI Compatible.
- Address (Base URL):
https://api.codexkey.online/v1. - Key (API key): your
sk_cdx_.... - Models (Models): a model ID such as
claude-sonnet-5.
Kilo can fetch the model list on its own when the address serves one. codexkey has GET /v1/models, but whether the dialog picks it up was not tested, so type the ID by hand if the list stays empty.
The dialog has no context size field. Kilo suggests keeping limits in kilo.jsonc, and the requirement is categorical: for custom and local models, always set limit.context and limit.output. The Local Config and Global Config buttons in the settings header open the file. The global one sits at ~/.config/kilo/kilo.jsonc, on Windows at C:\Users\<username>\.config\kilo\kilo.jsonc, and the project one lives in the repository root or in .kilo/kilo.jsonc.
On an older build whose settings show an API Provider dropdown, like Cline's, the fields are filled as in step 2. That is the screen the guide describes.
Context size: why the field is mandatory
How an extension behaves without a context size depends on the extension. Kilo, per its documentation, turns off automatic context management without limit.context and limit.output. The others fall back to a default, and the guide warns that the model may be treated as 128K. claude-sonnet-5 has a 1M context, so set it explicitly.
Still, you cannot copy 1000000 onto every model. In the Premium channel, for example, claude-haiku-4.5 has a 200K context and needs 200000. The numbers for each model are in the models table: send Max input to context and Max output to output.
Which model to pick
| Task | Model | Why |
|---|---|---|
| Everyday edits, tests, refactoring inside a module | claude-sonnet-5 | A sensible default for code, as in the guide |
| Hard refactoring, a plan for a large codebase | claude-opus-5-5 | The top Claude model, switchable in the same extension without changing the key |
| A GPT model in the same extension | gpt-6-sol | A working option from the catalog, test it on your own task |
Start with Sonnet and bring in Opus only when Sonnet stumbles on a specific task. How to test that without overspending is explained in Picking the model for the task without overpaying. Take exact IDs from the catalog: a typo in the name ends in Model not found.
How agent modes spend tokens
An agent works in a loop: it reads files, proposes an edit, runs a command and reads the result. Each step sends the accumulated history to the API as well, so a long task costs noticeably more than a short one. Two habits move the bill the most.
Do not turn on auto-approve right away. Per Cline's README, every file edit and terminal command waits for your consent until you switch auto-approve on, and while you watch each step, an agent stuck in a loop is visible at once, not on the bill in the evening. And start a new task when the old one has grown, instead of one conversation stretched across half a day. The other techniques, such as matching the model to the job and capping the reply length, are collected in How to cut your Claude Code bill threefold, and they apply to agent extensions the same way.
Which one to choose
For plugging in your own key, the choice between Cline and Kilo Code changes almost nothing: both want an address, a key and a model ID. Decide by the rest. Want Plan and Act modes and a desktop app? Cline. Want four agents, a cloud run in the browser and sign-in to an account with hundreds of models? Kilo Code. Roo Code is not worth taking into a new project: it is shut down.
Work in Cursor or Zed? You do not need these extensions, since each editor sets its address its own way: see Cursor and Zed with Continue. For Claude in a terminal and in VS Code, Anthropic has an extension of its own, covered in Claude Code in VS Code.
Common errors
As a rule, 401 means a stray space when pasting the key, a key cut off in copying, or a prefix other than sk_cdx_. Copy the key from the dashboard again.
Silence with no error: look at the /v1 in the address first. OpenAI Compatible mode requires it, Anthropic mode must not have it.
Model not found: compare the ID with the catalog character by character. There should be no spaces and no [1m].
History gets cut too early: the context size is not set or is left over from another model.
Tool-call errors. Roo's docs require the model to call tools in the OpenAI format, and providers with a partial implementation are where such errors come from. For Claude and GPT from the catalog this is usually no obstacle, but run obscure models on a tiny example first.
FAQ
Is Roo Code still alive?
No, the project is shut down. The README reports a closure on 15 May 2026, the repository is archived, and Cline and ZooCode are offered as replacements. Whether an already installed copy keeps answering was not checked.
What is the difference between Kilo Code and Cline?
The licenses differ: Kilo is MIT and Cline is Apache 2.0 (per the READMEs). Cline offers Plan and Act modes, Kilo offers Code, Plan, Ask, Debug and custom agents. Kilo also runs as a cloud agent in the browser, and it grants model access through its account or your key. For your own key both need the same thing: address, key, model ID.
How do I connect my own API key to Kilo Code?
Open settings, the Providers tab, then Custom provider at the bottom. Choose OpenAI Compatible as the API type, put https://api.codexkey.online/v1 in Base URL, then add the key and a model ID. Write the context and output limits into kilo.jsonc.
Does Kilo Code work in JetBrains IDEs like PyCharm and IntelliJ IDEA?
The plugin exists in the JetBrains Marketplace, and the README recommends searching for it under Settings, Plugins in any JetBrains IDE. How custom endpoint entry looks inside the plugin was not checked.
Is Kilo Code free?
The code is open under MIT, and the README promises provider pricing with no markup and sign-in without keys. The limits on free models were not checked. On your own key, tokens are paid to whoever issued the key.
Should the base URL end in /v1?
OpenAI Compatible mode needs it: https://api.codexkey.online/v1. Anthropic mode does not: https://api.codexkey.online.