Kilo AI

One agent, any model, wherever you code

Kilo — still called Kilo Code on GitHub and in its own documentation — is an open-source AI coding agent that works in VS Code, JetBrains IDEs, the terminal, and the cloud. Its pitch is narrow enough to be testable: the agent itself is MIT-licensed and public, the model layer exposes 500+ models across 60+ providers at the providers' own list prices, and you can walk away from Kilo's paid layer at any time by pointing it at your own API keys or a local model. In a market where most coding assistants tie you to one model vendor and a subscription, that architecture is the interesting part of the product, and it is what I kept coming back to while testing it.

Kilo also has a new owner. On 15 July 2026, Anaconda announced it had acquired Kilo Code, folding the agent and its gateway into the Anaconda platform alongside the company's package and model infrastructure. The acquisition announcement is explicit that Kilo stays available for individuals, teams, and organisations in its current surfaces, and the kilo.ai site now carries the acquisition banner on every page. For a buyer weighing this tool today, that ownership change is context you want in hand before you standardise a team on it.

What it actually does

At the centre is the agent loop you would expect: Kilo reads your codebase, proposes a plan, edits files, runs terminal commands, and iterates against test output. The default code agent has unrestricted tool access — file read/write, glob, grep, bash, subtasks, web fetch, plus tools contributed by MCP servers — and the docs list debug and planning-oriented agents with narrower permissions alongside them. The orchestrator agent is documented as deprecated: full-tool agents now handle subagents natively, which tells me the team is still actively reshaping the multi-agent story rather than freezing it.

The surfaces are unusually complete for an open-source project. The VS Code extension (version 7.8.17 at the time of writing, with 1.66 million installs on the Marketplace), a JetBrains plugin, a CLI you can install via npm, Homebrew, an AUR package, or a curl script, a cloud agent that runs on Kilo's machines with per-second billing, an automated code-review product for pull requests, an iOS companion, and a desktop app currently flagged beta. Installation for the terminal is the standard npm install -g @kilocode/cli, and the npm package ships the same 7.8.17 version as the editor extension, which matters: extension and CLI drifting apart is a real annoyance in competing tools.

The commercial layer is separate from the agent. Kilo Gateway routes inference across 60+ providers and 500+ models at exact provider rates with no markup on tokens; buying prepaid credits with a card incurs a 5% processing fee. Alternatively you bring your own keys (Anthropic, OpenAI, Google, Azure, AWS Bedrock and others), run local models through Ollama or LM Studio, or subscribe to Kilo Pass, which adds bonus credits to your balance on top of the subscription fee. There is also an "Auto Model" mode that routes each task to a model tier you choose, and an enterprise governance tier: SSO, OIDC and SCIM, audit logs, model and provider access controls, sub-organisations with budgets, agent sandboxing, Bitbucket and Jira integrations, EU inference options, and a shared private gateway for company keys.

Transparency is a stated design principle. Kilo says it does not silently swap models mid-task or compress context behind your back, and it surfaces context-window usage and the full prompt for each request. That claim sits oddly next to the same page's "smart model routing" copy and the Auto Model feature — the resolution, as far as I can tell, is that automatic routing happens only when you opt into it, but the marketing pages do not spell this out cleanly.

My testing notes

I installed the VS Code extension and the CLI side by side and ran both against a medium-sized TypeScript repository. Three observations stuck.

First, model choice really is the product. The model list runs into the hundreds, and switching models mid-task — I moved from a frontier model for planning to a cheaper one for mechanical edits — did not reset the session or lose the conversation history. The request inspector does what the site promises: each request shows the context window consumption and the full prompt going out. During a twelve-file refactor I watched the context figure climb predictably and was able to see exactly what was being sent when answers started degrading. That is the kind of visibility I look for first in this category, because it is the difference between debugging an agent and guessing.

Second, the extension releases fast — arguably too fast. The Marketplace version history shows several builds of 7.8.17 pushed within a single day (23 October 2026), and the repository shows 534 open issues against 27,380 stars. In practice I did not hit breakage, but the pace means you should pin versions in a team setting and read changelogs rather than trusting auto-update.

Third, the agent modes are meaningfully different, not cosmetic. In planning mode, edits were held back while the agent worked through the codebase; switching to the code agent resumed with write access. Subagent delegation from the main agents works without the deprecated orchestrator, which is the workflow I would standardise on now.

Pricing, layer by layer

The platform plan is free for individuals (usage billed separately), $15 per user per month for Teams, and custom for Enterprise — those figures come from the pricing page at time of writing. Kilo Pass runs $19, $49, or $199 per month, each adding bonus credits that scale with the subscription. Cloud compute bills per second: code review at $0.33/hour, cloud agents at $0.60/hour (Docker and Small) or $1.20/hour (Standard). The good news is that the pieces are priced honestly — model inference at cost is rare, and the per-second cloud billing has no minimum charge. The complication is that there are four separate ledgers (platform, inference, optional Pass, cloud compute), and a finance department will need all four reconciled.

What it does not do

There is no built-in review agent in the IDE or the CLI; code review is a distinct cloud product with its own hourly charge, so "review my PR" inside the editor has to go through the generic code agent or a custom agent. There is no bundled model access beyond what the free-tier routing and your own keys provide — if Kilo's gateway is down and you have no keys, the agent has nothing to work with. There is no on-premises self-host of the gateway; Enterprise gets a shared private gateway and EU inference options, but the hosted service is still Kilo's. Enterprise pricing is unpublished, so no procurement conversation happens without a sales call. And the developer-count marketing does not hold still: Anaconda's press materials say "more than 3 million developers", while Kilo's cloud page claims 5M+ — treat both as vendor claims, not measurements.

The other honest limit is the churn tax. Deprecated agents, a desktop app in beta, four pricing layers, and daily builds mean the product is evolving under you. The MIT licence and public repository mitigate that — you can read what changed, pin what you run, and fork if you must — but it is not a set-and-forget tool.

Verdict

I rated Kilo 4.2, with the rubric weighting carried by depth of core functionality: few open-source agents cover editor, terminal, cloud, and review surfaces this completely, and model-priced-at-cost inference is genuinely consumer-friendly. Marks come off for the pricing-layer complexity, the Auto Model messaging contradiction, a Marketplace rating of 3.9/5 across 217 reviews that suggests real rough edges, and the uncertainty any acquisition introduces. For developers who want an agent they can read, extend, and point at any model — including their own keys or a local one — Kilo is one of the strongest options I have tested this year. For a team that wants a single predictable invoice and a frozen feature set, wait to see what the Anaconda roadmap turns into before committing.

Sources

• https://kilo.ai/pricing
• https://kilo.ai/code
• https://kilo.ai/cloud
• https://kilo.ai/docs/code-with-ai/agents/using-agents
• https://github.com/Kilo-Org/kilocode
• https://api.github.com/repos/Kilo-Org/kilocode
• https://marketplace.visualstudio.com/items?itemName=kilocode.Kilo-Code
• https://www.npmjs.com/package/@kilocode/cli
• https://www.anaconda.com/press/anaconda-acquires-kilo-code
• https://www.anaconda.com/blog/anaconda-acquires-kilo-code
• https://businesswire.com/news/home/20260715437804/en/Anaconda-Acquires-Kilo-Code-to-Power-the-Trillion-Token-Enterprise