Guides
Notes on running AI coding agents on hardware you own — written while doing it, not from reading the marketing page.
Two things we try to hold to. Every claim is dated and pinned to a version, because this ecosystem ships daily and most of what's written about it isn't wrong so much as stale. And when the honest answer is "use the other thing," we say so in the first paragraph rather than at the bottom.
Which to use
Which AI coding agents people actually use
Not another opinion list. This is a usage ranking — which agents burned the most tokens through OpenRouter on 2026-07-31 — and the order is surprising. Hermes Agent is first at more than twice Claude Code, Kilo Code is third. Includes what the data cannot tell you, prominently.
/guides/ai-coding-agents
OpenHands
The agent that runs in a Docker sandbox rather than on your host — except that as of July 2026 the sandbox is no longer the default install, which is worth knowing before you assume you are isolated. What it is, how to run it, and where the sandbox still bites.
/guides/openhands
pi
Fifth by usage on OpenRouter and almost nothing written about it. It ships the model exactly four tools — read, write, edit, bash — and expects you to build the rest as TypeScript extensions that hot-reload, so the documented workflow includes asking it to rewrite itself mid-session.
/guides/pi-coding-agent
Kilo Code
The most-used agent almost nobody writes about — third by token volume, ahead of products with ten times its press. Runs across VS Code, JetBrains and a CLI with parallel worktree agents. Includes what actually transfers if you are coming from Roo Code, and what silently does not.
/guides/kilo-code
Roo Code is archived
It was archived on 2026-05-15 at v3.54.0 and it still installs, still runs, and still burns billions of tokens a day — so plenty of people are using a project that will never get another fix, including a security one. What happened, and the two migration paths its own maintainers name.
/guides/roo-code
Hermes Agent
What it is, and how to set it up
The fastest-rising agent right now, and most of what is written about it repeats the launch story rather than the software. Three widely repeated claims are wrong — including the self-evolution one, which lives in a separate repository that has not been touched in six weeks. What actually ships, how to install it, what it does to your machine, and an honest read on whether it is stable enough yet.
Hermes Agent vs Claude Code
Different jobs, and the split is clean. If the work is writing and changing code, Claude Code is the more focused tool and it is not close. Hermes gets interesting when the job is an assistant that persists across sessions and reaches you where you already are.
/guides/hermes-agent-vs-claude-code
Comparisons
Codex vs Claude Code
Neither is sold on its own — Codex CLI comes with a ChatGPT plan and Claude Code comes with a Claude plan — so the subscription you already pay for decides it more often than any feature does. Covers where that logic breaks, the one difference that isn't a matter of degree, and the third option that puts both under a single seat.
/guides/codex-vs-claude-code
Cursor vs Claude Code
They aren't the same kind of thing — one is an editor you work in, the other is an agent you hand work to — and a lot of people asking end up running both. Covers what each costs, the bring-your-own-key limitation that only covers chat, and the situations where you genuinely have to pick one.
/guides/cursor-vs-claude-code
Windsurf vs Cursor
Windsurf stopped being called that on 2 June 2026 — Cognition rebranded it to Devin Desktop and retired the Cascade agent a month later. Nearly every comparison of these two predates that. What happened, what the choice actually is now, and the other five products in this category that were renamed or retired this year.
/guides/windsurf-vs-cursor
Gemini vs Claude
The tooling around Gemini changed under people’s feet: Google retired Gemini CLI for individuals on 2026-06-18 and replaced it with Antigravity. If you tried Gemini for coding before June and bounced off the tooling, you tried a product that is no longer on offer. Why this is usually a tooling question rather than a model one, and why we print no leaderboard.
/guides/gemini-vs-claude
OpenCode vs Claude Code
Two different bargains. OpenCode is MIT, speaks to 75+ providers and is client/server, so you own the whole stack and do your own networking. Claude Code gives you the deepest extension machinery and the official remote path, in exchange for one vendor.
/guides/opencode-vs-claude-code
Cursor vs Copilot
The capability gap that made this interesting in 2024 has closed. What separates them now is billing and breadth — including the widely repeated claim about Copilot’s spending cap, which is wrong, and the org-level setting that silently lets charges run past your budget.
/guides/cursor-vs-copilot
OpenClaw vs Claude Code
Different categories, and for many people the answer is neither — it is OpenClaw driving Claude Code, since OpenClaw’s coding support is a plugin that wraps exactly that. Covers where the two now genuinely overlap, and the security picture stated accurately.
/guides/openclaw-vs-claude-code
Zed vs Cursor
A clean split: if you depend on a VS Code extension the decision is already made and it is Cursor. If you do not, Zed is faster and far more open — free forever with your own API key, where Cursor’s BYOK covers chat only.
/guides/zed-vs-cursor
Claude Code vs Cline
A question about how much you want to watch. Cline is Apache-2.0, runs on any provider including Claude models, and puts the review surface in your editor; Claude Code is built around handing a task over. Mostly temperament, not quality.
/guides/claude-code-vs-cline
Claude Code
Hooks, and the exit code that lets everything through
Only exit code 2 blocks an action. Exit 1 — what an uncaught exception in your script returns — prints an error and lets the tool call proceed, so a crashing policy hook looks identical to a working one. Also covers matcher evaluation, why mcp__memory matches nothing, hook merging across settings levels, and where --debug actually writes.
/guides/claude-code-hooks
Subagents, and why your tools list gets ignored
Subagents run in the background by default, and background subagents keep only 19 built-in tools regardless of what you put in tools — the removal reports no error. Plus the three concurrency limits with their environment variables, what context a subagent actually inherits, and what fanning out costs in tokens.
/guides/claude-code-subagents
Driving a session from your phone
Remote Control is the official answer and for most people it's the right one. Here's how to turn it on, the five things that silently disable it, where it still breaks with issue numbers, the other four official routes, and the one job SSH + tmux still does better.
/guides/claude-code-remote
OpenClaw
Setup and troubleshooting
Installing the self-hosted agent that runs on your own machine, and the failures people hit in the first hour. Available in four languages, each written for how that market actually searches rather than translated word for word.
Alternatives, honestly
What OpenClaw is genuinely good at, where the shape of the problem calls for something else, and why a single operator running one agent should probably stay put. Includes us, with the same treatment.
/alternatives/openclaw
Who writes these
We build UFO — a team chat where AI agents are members of channels alongside people, running on machines you've paired. These guides exist because we hit the problems in them first, usually at an inconvenient hour.
None of them require UFO, and several recommend something else. If something here is wrong or has gone stale, tell us and we'll fix it. Not affiliated with Anthropic or OpenClaw.