YAI Setups Review · 2026 edition Jan 1 – Nov 5, 2026 Hacker News · Lobste.rs · Tildes · blogs

Infographic report · counted by unique authors

What the IT community actually uses for AI work

I wanted to know what people actually run, not what gets announced. So I went through comments from Hacker News, Lobste.rs and Tildes posted in 2026, plus a pile of personal blogs. Each person counts once per tool, and only if they say they use it now. This edition also looks at the interface people work in and at how the three communities differ.

of authors who described their setup use Claude Code.

Harnesses: Claude Code, then a long gap

Agents and coding tools ranked by the number of authors who currently use them. Percentages are out of authors who named at least one tool.

Top 18 · harness / coding tool

Each dot = 1 author

Plenty of people run both. authors use Claude Code and Codex CLI side by side, usually Claude Code for the main work and Codex for review or when Claude gets stuck.

Beyond coding · apps and services

Communities side by side: same leader, different tail

Share of authors in each source who use the tool. Cell colour reflects the share within the row. Blogs means authors of articles about their own setup, one vote each. People who blog about their AI setup tend to like AI, so expect higher numbers in that row.

Interfaces: mostly the terminal

Where people actually type: terminal, IDE, desktop app, browser, phone. Solid bars are people who said so. Hatched bars are my inference from the tool's default interface (someone who says "Codex CLI" is in a terminal), because most people never mention it.

Interface type · unique authors · explicit inferred

Host: where it runs (explicitly stated)

Claude Code: interface where the author named it

GitHub Copilot: interface where the author named it

Model × what the author uses it through · number of authors who stated the pairing explicitly

Claude mostly goes through Claude Code. Open models like Qwen, DeepSeek and GLM go through Pi and OpenCode, through OpenRouter, or straight into llama.cpp. GPT usually shows up with Codex CLI. No surprises there, but it is nice to see it in numbers.

Trend: open-source agents up, Cursor down

Share of authors in each quarter who use the tool. I use shares, not counts, because sample size jumps around between quarters. Q4 is mostly one big HN thread from early November. All panels share the same axis, and the grey lines are the other tools.

Models: Claude first, open weights not far behind

Solid bars are people who named the model. Hatched bars are inferred from a harness that only talks to one provider: Claude Code means Claude, Codex and ChatGPT mean GPT, Gemini CLI and Antigravity mean Gemini.

Model groups · explicit implicit open weights

Families · explicit mentions only · most common versions

A lot of the Qwen and DeepSeek mentions come from a single June thread, "Has anyone replaced Claude/GPT with a local model…". Take that thread out and Qwen is still the open model people name most.

Money: median a month, a tail in the thousands

Who pays whom, and how much, for the people who mentioned it. Most don't, so the spend chart is a small sample.

Each dot = 1 author with a stated amount · log axis USD/mo

Breakdown by band

<$20$20–100$100–250>$250

Plans and billing modes · top 18

Churn: who left what, and for what

Authors who wrote that they stopped using a tool ( author × tool cases in total, across all categories). Left is what they left, right is where they went. I only draw moves with a known destination. The rest are counted next to the tool as "no destination".

Claude Code and Codex CLI trade users in both directions. That reads to me less like loyalty and more like people following whichever model is better this month. Cursor and Aider users who left went mostly to Claude Code. OpenClaw lost people after Anthropic changed its subscription rules, and some of them ended up on Hermes Agent.

Local models: llama.cpp ahead of Ollama

authors mention running their own local inference (runtime or hardware).

Runtime

Hardware (classes)

Macs with a lot of unified memory are the most common box, then gaming GeForce cards from the 3090 to the 5090. The 128 GB DGX Spark / GB10 and AMD Strix Halo machines come up more often than I expected.

Calibration: these communities vs. the big 2026 surveys

Community share = authors using the tool / authors who named at least one coding harness. Each survey asks a different question ("in the past year", "weekly at work", "regular agent users"), so compare the order and the direction, not the exact percentages.

These three communities lean hard toward the terminal and open source. Pi gets 3× its Stack Overflow share here, OpenCode 2× its JetBrains share. Copilot, Cursor and JetBrains AI sit well below the surveys. My guess is that their users are mostly people at companies who simply don't post in threads like these, but I can't prove that from this data.

Sources: Stack Overflow Developer Survey 2026 (n≈12,260, "used in the past year") · JetBrains Developer Ecosystem 2026 (n>15,000, May–Jul, "at work") · Pragmatic Engineer AI Tooling 2026 (n=507 regular agent users).

How the bloggers do it

articles about personal setups from 2026 by authors (HN front page + library and ITech feeds in Readwise Reader, including Swamp Club). Only the author of each article counts, not the comments under it. Below: one article per author, up to 24.

Sources and the richest threads

How much I read in each place, how many people described a setup, and the threads where the most people did.

Methodology and limitations

Process

I picked 21 Ask HN threads about setups and workflows by hand (19 had 2026 comments) and 213 Lobste.rs and Tildes discussions about using AI in practice.
I downloaded the full comment trees and kept only comments from 2026 ().
An LLM read every one of them, no keyword pre-filter, and pulled out the tool, whether the person still uses it, the interface, the host and which model runs in which harness.
Blogs came from the HN front page and my own Readwise Reader library and feeds (including Swamp Club). I triaged them by title and summary, then an LLM read the full text. Only the author counts.
Tool names were normalised, duplicates removed per author and tool, and the latest statement wins.

Counting rules

One person, one tool, one vote per site. The same handle on two sites counts twice, because I can't tell if it is the same person.
Only the author's own current use counts, not recommendations or "I've heard".
Interfaces are shown two ways: what people said, and what I inferred from the tool's default.
The surveys are only a reference point. They are not mixed into the rankings.
Tools with fewer than 3 users are left out of the charts.

Biases and limitations

People who comment on HN, Lobste.rs and Tildes are not typical developers. They lean toward the terminal and open source.
One June HN thread about local models pushes up Qwen, DeepSeek, Pi and OpenCode.
Swamp is the product behind the Swamp Club feed, so I left it out of the comparison between sources.
No Reddit. Its Responsible Builder Policy requires approval for API access and does not allow research outside the Reddit for Researchers program.
The LLM can misread things. As a check, 99.6% of the extracted quotes match the source comment word for word.