We read comments from Hacker News, Lobste.rs and Tildes posted in 2026, plus personal blogs. Each author counts once per tool, and only if they currently use it themselves. New in this edition: which interface people use to work with AI and how the communities differ.
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
Claude Code and Codex CLI are often not mutually exclusive. authors use both, typically Claude Code as the main tool and Codex for review or as a fallback.
Beyond coding · apps and services
Share of authors in each source who use the tool. Cell colour reflects the share within the row. Blogs = authors of articles about their own setup (1 vote per author). They are mostly enthusiasts, so naturally “pro-AI”.
Which surface people use to work with a tool or model. Solid fill = the author named the interface explicitly. Hatching = inferred from the tool’s default interface (e.g. “Codex CLI” → terminal), since most people don’t 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 is used mostly through Claude Code. Open models (Qwen, DeepSeek, GLM) run mainly through the open harnesses Pi and OpenCode, via OpenRouter, or directly on llama.cpp. GPT is most often paired with Codex CLI.
Share of authors in each quarter who use the tool. Shares rather than absolute counts, because sample size varies (Q4 is mostly one big HN thread from November). Every panel uses the same axis; grey lines are the other tools for comparison.
Solid fill means the author names the model explicitly. Hatching means the model is inferred from a harness tied to a single provider (Claude Code → Claude, Codex/ChatGPT → GPT, Gemini CLI/Antigravity → Gemini).
Model groups · explicit implicit open weights
Families · explicit mentions only · most common versions
Qwen and DeepSeek owe much of their count to the thread “Has anyone replaced Claude/GPT with a local model…”. Even without it, Qwen remains the most frequently named open model.
Paid plans by provider (unique authors) and monthly spend for those who stated it.
Each dot = 1 author with a stated amount · log axis USD/mo
Breakdown by band
Plans and billing modes · top 18
Authors who wrote that they stopped using a tool ( author × tool cases in total, across all categories). Left is the original harness or app, right is the destination. Only departures with a known destination are drawn; the rest are listed next to the source as “no destination”.
People move between Claude Code and Codex CLI in both directions, so it looks more like switching to whichever model is better at the moment. Cursor and Aider lose almost exclusively to Claude Code. OpenClaw lost users after Anthropic changed its subscription rules, partly to Hermes Agent.
authors mention running their own local inference (runtime or hardware).
Runtime
Hardware (classes)
Most common are Apple Silicon Macs with lots of unified memory, followed by gaming GeForce cards (RTX 3090 to 5090). 128 GB DGX Spark / GB10 and AMD Strix Halo boxes also show up fairly often.
Community share = authors using the tool / authors who named at least one coding harness. The surveys ask differently (“in the past year”, “weekly at work”, “regular agent users”), so compare rank and direction, not absolute percentages.
HN, Lobste.rs and Tildes skew strongly terminal and open source compared with the average: Pi’s share in the community is 3× its Stack Overflow figure, and OpenCode’s is 2× its JetBrains figure. Copilot, Cursor and JetBrains AI, on the other hand, are well below average. Those tools are used mainly by the “silent majority” at companies, who don’t post in threads like these.
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).
articles about personal setups from 2026 by authors (HN front page + library and ITech feeds in Readwise Reader, including Swamp Club). Only the article’s author was counted, not the comments under it. Below is one article per author, at most 24.
How much we read, how many authors described their own setup, and which threads were richest (by number of authors with a setup).