UFO

Kilo Code

Short answer: it is a genuinely mainstream tool that reads as a niche one, and the reason to pick it is breadth rather than any single standout feature. Kilo Code runs in VS Code, in JetBrains IDEs and as a terminal CLI, with parallel worktree-isolated agents, subagent delegation and a very large model catalogue. Nothing on that list is unique. Having all of it under one MIT licence, with no vendor lock on which model you use, mostly is.

The odd part is how little gets written about it. On OpenRouter's coding leaderboard it sits third, ahead of several products with ten times its press. This page is what we could verify about it on 2026-07-31.

Written 2026-07-31 · Checked against kilocode v7.4.17 (released 2026-07-29), JetBrains plugin 7.0.11, and the docs at kilo.ai · Star counts and version numbers in this category go stale in weeks

On this page

  1. The number that doesn't match the profile
  2. What it is, and where it runs
  3. The Roo Code inheritance
  4. Agent Manager and subagents
  5. Models, BYOK, and the free path
  6. Installing it
  7. What it costs
  8. Where it's rough

The number that doesn't match the profile

OpenRouter publishes a leaderboard of the applications routing the most tokens through it. Reading the coding category on 2026-07-31, Kilo Code was third for that day at 241B tokens, behind Hermes Agent and Claude Code and ahead of OpenClaw, Cline, Codex and Cursor. Its all-time figure was 7.2T, which put it marginally ahead of Claude Code's 7.09T on the same board.

That number needs a caveat, and it is a big one. OpenRouter only counts applications that opt into usage tracking, and only traffic that goes through OpenRouter at all. Anything running on a vendor subscription or a direct API key is invisible to it. Claude Code and Codex are certainly under-counted here, so the board is not a market-share ranking.

What it does measure is real work by users who chose their own routing, and on that narrower question Kilo Code is near the top. Set against 26,639 GitHub stars, 1,370,833 VS Code Marketplace installs and 325,197 JetBrains Marketplace downloads — all read on 2026-07-31 — the picture is consistent: a lot of people use this, and very few write about it.

What it is, and where it runs

Kilo Code is an open-source coding agent, MIT-licensed, from Kilo-Org. The same agent is packaged for four places:

The important structural fact is that these are not three separate implementations. The VS Code extension was rebuilt on the CLI's core, which its release notes describe as "a portable, open-source core shared across VS Code, the CLI, and Cloud Agents." That core is OpenCode: the monorepo carries a packages/opencode directory and the top-level dev script runs out of it. If you have read our OpenCode comparison, the client/server split described there is the machinery underneath Kilo too.

Out of the box you get five primary agents you switch between — Code, Plan, Ask, Debug and Review — plus inline autocomplete, terminal and browser control, MCP server support, and an autonomous kilo run --auto mode intended for CI. The docs are explicit that --auto disables every permission prompt, so it is a trusted-environment flag, not a convenience one.

The Roo Code inheritance

Part of why the usage number is high is that Kilo absorbed another project's users. Kilo's own migration guide is unusually direct about the lineage:

We started Kilo as a Roo fork in 2025. The two codebases share real git history.

Roo Code shut down. Its final release was v3.54.0 on 2026-05-15, the repository was archived the same day, and the team moved to a cloud product. Roo had reached three million installs, and Kilo is one of the two places those users were pointed. We cover the archived project separately in the Roo Code guide.

If you are migrating, the guide is worth reading in full rather than trusting the wizard, because the honest parts are buried:

If you depended on Orchestrator, Memory Bank or Qdrant-backed indexing, Kilo's escape hatch is to pin the legacy v5.x line, which still ships the Roo-derived codebase with those features intact. That is a real option, and also a version you are choosing to freeze — treat it as a deadline, not a destination.

Licence footnote, since people ask: Roo Code was Apache-2.0; the Kilo repository declares MIT.

Agent Manager and subagents

Two features get conflated in write-ups. They sit at different layers.

Agent Manager is a UI. It opens as a full-panel editor tab in VS Code (Cmd+Shift+M / Ctrl+Shift+M) and runs several sessions side by side, each in its own git worktree on its own branch, created under .kilo/worktrees/. It carries a diff panel against the parent branch, a dedicated terminal per session, setup scripts and .env auto-copy on worktree creation, and session import from an existing branch or a GitHub PR URL. PR status badges require the GitHub CLI to be installed and authenticated; without gh they just don't appear.

The docs flag a cost that is easy to underestimate: worktrees share git objects but are separate checkouts, so node_modules, build output and caches multiply across parallel agents, and closing one does not clean up containers, volumes or databases your setup scripts created outside it.

Subagents are a delegation mechanism, mostly a CLI concern. Two ship built in — general and a read-only explore — and you define your own in kilo.jsonc or as markdown agent files, each with its own prompt, model, tool access and permissions. A primary agent invokes them through the Task tool, or you call one directly with @agent-name. Each runs in an isolated session and returns a summary to its parent. There is no UI for configuring them yet, and subagents cannot spawn subagents.

Review is the third piece. Beyond the Review agent, /review runs a local AI review with line-level inline comments — bare, it covers staged, unstaged and untracked changes, and it also takes uncommitted, branch [base], a commit hash, or a PR number.

Models, BYOK, and the free path

Kilo advertises 500+ models across 60+ providers with mid-task switching, and the provider docs list the expected set — Anthropic, OpenAI, Google, Bedrock, Vertex, Azure, Mistral, DeepSeek, Groq, xAI, Moonshot, OpenRouter — plus local models through Ollama and LM Studio, and a generic OpenAI-compatible option.

BYOK is supported properly, not partially: you can point it at your own provider keys on the free tier, which is not true of every tool in this category. There is also a hosted route, the Kilo Gateway, whose pitch is "Pay exact provider rates with no markup."

A genuinely free path exists, but it has three parts and the docs are clear that you must configure all three. Models are used for agentic work, for autocomplete, and for background tasks (session titles, context summarisation) — leave any one on a paid model and you are still spending. The easy option is Auto Free (kilo-auto/free), which routes to whatever free models are available.

Read the Auto Free data-handling warning before you point it at work code. Kilo's own documentation states that Auto Free "may route your requests to providers that log prompts and outputs and use them to improve their services," naming NVIDIA's free endpoints specifically, and says: "Do not submit personal or confidential data when using Auto Free." Free here means the provider is getting something else.

Installing it

CLI — pick one:

npm install -g @kilocode/cli
curl -fsSL https://kilo.ai/cli/install | bash
brew install Kilo-Org/tap/kilo

pnpm, bun and the AUR work too, and the releases page carries prebuilt binaries for Windows, macOS and Linux — including a x64-baseline build for older CPUs without AVX and a musl build for Alpine. Then run kilo in a project directory.

VS Code — install kilocode.Kilo-Code from the Marketplace, or:

code --install-extension kilocode.Kilo-Code

JetBrains — search "Kilo Code" in Settings → Plugins, or install plugin 28350 from the Marketplace. Note the JetBrains build tracks a separate version line and lags the VS Code one.

An account gets you the hosted catalogue; with your own key or local models you don't need one.

What it costs

As of 2026-07-31, from the vendor pricing page. The platform and the inference are billed separately, which is the part people misread.

PlatformPriceWhat it adds
Free & Open Source$0Everything in the repo. Usage billed separately
Teams$15/user/monthShared BYOK, usage analytics, shared agent modes, centralised billing. 14-day trial
EnterpriseCustomSSO/OIDC/SCIM, audit logs, model access controls, shared private gateway, SLA

For inference you then choose: $0 with BYOK, local models or Auto Free; Kilo Gateway at usage cost billed at provider rates; or Kilo Pass, a prepaid credit scheme starting at $19/month with higher tiers at $49 and $199. A 5% processing fee applies to credit purchases. Cloud compute — cloud agents, hosted code review — is metered separately by the hour and billed per second.

The short version: the tool is free and the models are the bill, and if you already hold provider keys the platform costs you nothing.

Where it's rough

All of these were open on the public tracker on 2026-07-31, against 694 open issues in total.

None of this is unusual for the field — Cline has its own long-running problems, and everything here ships weekly. It is worth knowing which specific things bite before you standardise a team on one.

Where we fit, briefly

We build UFO, so weigh this accordingly. If the question is "should I use Kilo Code," the answer on this page is not us — it is Kilo Code, and it is free to try in the editor you already have.

UFO is the layer after: a team chat where AI agents are members of channels alongside people, each running on a machine you've paired, with task cards for work that outlives a session. Kilo Code is a thing you would run inside that. The question it answers starts once there are several agents, on several models, and other people who need to see what they did.

Sources

Repository metadata, release versions and issue states were read from the GitHub API on 2026-07-31. Marketplace counts were read from the VS Code and JetBrains marketplace APIs the same day. Pricing and feature claims come from Kilo's own documentation. The OpenRouter figures are from our own collection on 2026-07-31 and carry the tracking-opt-in caveat described above.

Not affiliated with Kilo-Org. If something here is wrong or has gone stale, tell us and we'll fix it.