In XDA Developers’ account, Parth Shah describes replacing those three tools with Kilo after building a fictional perfume-store website. He divided the homepage, product-browsing experience, and shopping cart across separate sessions, then experimented with different models for design changes and debugging.
That is a useful workflow example. It is not a controlled comparison: the account supplies no matched benchmark, project-completion timings, cost breakdown, or test results demonstrating that Kilo outperformed all three competitors.
The more interesting question is therefore practical: does Kilo reduce tool-switching without creating more work in model selection, review, and administration?
Model choice is the attraction—not proof of better code
Kilo advertises access to more than 500 models. That is a vendor-reported catalog figure, not an independently audited count or a promise that every model is equally suitable for coding. Its desktop documentation confirms support for hosted models through Kilo Gateway, users’ own provider API keys, and local inference.
For a developer, the potential benefit is straightforward: try a different model without moving the project into another coding environment. Shah reports using that flexibility to match models to different stages of his website project.
There is also an important architectural detail behind the desktop interface. According to Kilo’s overview, the application starts a local copy of Kilo CLI and uses the same runtime that powers the CLI and VS Code extension. The desktop app is another interface to that agent environment—not evidence of a separate, more capable coding engine.
The takeaway: Kilo offers flexibility in choosing where inference happens. Whether that produces better software remains something to measure on the actual project.
Auto routing has a privacy boundary worth reading
Kilo’s Auto Model feature selects an underlying model for requests. Its documented choices are:
- Frontier: paid routing focused on capability.
- Efficient: paid routing intended to match task difficulty with cost.
- Free: routing among available free models.
Kilo says the underlying models can change server-side as availability and pricing change. The expanded model picker displays the model used and its cost. Those are useful controls, but the routing descriptions should not be mistaken for independently demonstrated savings.
The significant caveat is Auto Free’s data handling. Kilo explicitly warns that requests may reach providers that log prompts and outputs or use them to improve their services. Its documentation tells users not to submit personal or confidential information through that tier.
For company repositories, that warning should come before the first prompt—not after the first invoice.
There is a version boundary, too: automatic mode-based switching requires VS Code or JetBrains extension version 5.2.3 or later, or CLI 1.0.15 or later. Older versions default to one model for requests. These are documented minimums, not claims about the latest release.
Parallel agents need explicit isolation
Shah’s perfume-store example illustrates the attraction of parallel sessions: several bounded tasks can progress separately instead of waiting in one conversation.
Kilo’s Agent Manager documentation confirms support for both isolated Git worktrees and sessions sharing a local workspace. Crucially, a delegated task subagent does not automatically create a worktree. Developers who need separate checkouts and branches must use Agent Manager’s worktree mode.
That distinction matters. “Multiple agents” does not necessarily mean “separate filesystems.”
For the documented VS Code workflow, Kilo requires an open workspace folder and a Git repository for worktree features. It also instructs users to open the main repository rather than an existing worktree checkout when creating new worktrees. Agent Manager includes a diff/review panel and dedicated terminals for sessions.
A sensible evaluation would therefore begin with clearly bounded tasks and explicit isolation, followed by review and integration testing. That is an engineering recommendation, not a claim of firsthand testing. Giving three agents keyboards does not appoint a release manager.
What Windows developers should check
Kilo’s installation page lists a desktop application for Windows, macOS, and Linux, alongside VS Code, JetBrains, terminal, and cloud access. The desktop offering is labeled beta. Windows installers are available for both x64 and ARM64 PCs.
Its installation documentation adds several practical requirements:
- Desktop Git integration uses system Git 2.17 or newer.
- The local model server, notebooks, and conda environments require signing into a Kilo account.
- Local inference needs a model file in GGUF format.
- On Windows, updates download in the background and require restarting the application to install.
Local inference deserves its own qualification. Kilo says its managed local server can process prompts without sending them to a hosted model provider. However, the server still requires a Kilo account, and downloading models from its catalog additionally requires an Anaconda account. Users can instead import a GGUF file already on disk. Local processing is not the same as account-free operation.
A workflow alternative, not an obituary
Shah’s conclusion is strongest when read as a personal preference: Kilo brought enough of his preferred capabilities together that he no longer wanted to juggle three tools. His account also acknowledges desktop bugs, potentially rising premium-model expenses, and the continuing need to supervise agents.
The evidence supports treating Kilo as a flexible alternative worth evaluating against concrete needs. It does not support calling Claude Code, Codex, or Google Antigravity outdated.
For developers considering a switch, the useful scorecard is not the number of models in the picker. It is accepted code, review effort, total spending, and whether the chosen data-handling arrangements are appropriate for the repository. Consolidation is valuable when it removes friction—not merely when it moves every moving part into one window.