Kilo Code is positioning itself as the agentic development environment for developers who want to hand off the grind work. Available as a VS Code extension, JetBrains plugin, CLI (via Homebrew), or cloud-hosted service, it lets you prompt an AI agent to plan, implement, and debug code with minimal manual intervention. One developer recently documented their experience using Kilo's free tier to build a production-grade Rust application from scratch—and the results reveal both the promise and the quirks of modern agentic coding tools.

Setting Up Kilo Code

Installation takes under two minutes: brew install Kilo-Org/tap/kilo, register for a free account at kilo.dev, then authenticate with kilo auth login. The author used the 'Auto Free Kilo Gateway' model throughout, which routes requests to available free endpoints—a critical detail that sets realistic expectations about response times and capability ceilings. After firing up kilo in the terminal, slash commands (/) trigger agent actions like /models for model selection.

The Project: Crypto Price Tracker

The use case was practical: a Rust CLI application calling five cryptocurrency APIs (CryptoCompare x2, OpenExchangeRates, Valr) to fetch real-time price data and output results to a LibreOffice Calc ODS spreadsheet. The author had been running an aging bash script via cron that produced unwieldy text files—Kilo Code became the vehicle for upgrading this into structured, queryable data storage with timestamps in YYYY-MM-DD HH:MM format.

How Kilo Handled Planning and Implementation

The agent operated in two distinct phases. First, it generated a detailed implementation plan when prompted with /plan mode—a breakdown of API calls, data transformations, and file handling logic. The author then refined the scope by requesting modifications: auto-create the ODS file if missing, append to existing files otherwise, and extract only .lastTradedPrice from Valr's response. After approval, switching to code generation mode kicked off implementation using Rust with the calamine crate for ODS handling—a dependency the agent initially got wrong before self-correcting after build failures.

Session Stats Reveal Token Appetite

The numbers are eye-opening: 115 requests made during the session, consuming 7,475,473 input tokens and producing 42,900 output tokens. This highlights a fundamental reality of agentic coding—the context window fills fast when you're iterating on a full application rather than snippets. The binary compiled successfully, runs in milliseconds as a cron job, and handles the defined scope cleanly.

Manual Tweaks Still Required

Despite impressive autonomy, human oversight remained necessary. UTC timestamps were switched to local time for South African timezone alignment (the author appears to be based in SA given Valr API usage), and println! output was commented out since cron jobs run headless without terminal access. These aren't failures—they're the seam between agent capability and production deployment realities.

Key Takeaways

  • Agentic coding tools like Kilo Code handle boilerplate, dependency research, and iterative debugging surprisingly well on free tier models
  • Token consumption scales rapidly with application complexity—budget awareness matters for sustained use
  • Self-correction during build failures demonstrates viable error-handling loops in modern AI coding agents
  • Human review remains essential for timezone handling, deployment context, and production edge cases

The Bottom Line

Kilo Code delivered on its core promise: turning a vague requirement into working, compiled Rust code using free models. That's not science fiction anymore—it's Tuesday afternoon workflow optimization. If you're still writing every line by hand, the question isn't whether agentic tools will change your practice, it's how long you want to wait around for them to do it.