GitHub Top-100 in 2026 — Agent Harnesses Ate the All-Time Chart

Snapshot taken 2026-07-29 from star-history.com (all-time ranking, 12 142 repos tracked; metadata verified against the GitHub API).

The headline is not “AI is popular.” It is that the unit of open-source distribution changed. Sixteen repos in the all-time top-100 were created after January 2025, and all sixteen are agent harness tooling — skills, config files, agent CLIs, provider switchers. Not one is a library, framework, or runtime.

The young cohort — 16 of 100

For scale: build-your-own-x at #1 took eight years to reach 533K (177★/day). React is at 247K after thirteen years (~52★/day). A markdown file with four rules for CLAUDE.md is pulling 20× React’s lifetime rate.

Overall composition of the top-100: 33 repos are AI/LLM-related, 17 of them are agent harness tooling specifically, and 16 of those 17 are younger than 18 months. Learning resources still hold #1–#5 and #8–#13; infrastructure (linux, vscode, react, node, kubernetes) is stable but no longer where the attention flows.

Why this matters for us

1. Markdown is now a shippable product. Five of the sixteen contain effectively zero application code — andrej-karpathy-skills is literally one CLAUDE.md. This is the strongest available evidence for harness engineering: the leverage is in the instructions, not the runtime. solo-factory (27 skills) is in exactly this category and is competing in a category that now has a top-100 ceiling.

2. gstack is the closest public analogue to solo-factory. Garry Tan packages 23 slash commands as roles — CEO, eng manager, designer, reviewer, QA lead, security officer (OWASP + STRIDE), release engineer. Ours are pipeline stages (/research → /validate → /plan → /build → /launch). Two different decompositions of the same problem: roles vs stages. Worth stealing: /office-hours (describe what you’re building, get product pushback before planning), /qa against a live staging URL with a real browser, and a security-officer pass as a gate rather than an afterthought. His productivity claim — 11.4K logical lines/day in 2026 vs 14/day in 2013, with a published reproduction script — is the first serious attempt to measure the solo-with-agents delta instead of asserting it.

3. “Taste skills” are a new category. In the same week’s trending list: ponytail (“makes your agent think like the laziest senior dev in the room”), hallmark (anti-AI-slop design skill), impeccable (“design language that makes your harness better at design”). These encode judgment, not procedure — the opposite of a checklist skill. Our /humanize is the same species. See agent-toolkit-landscape for the full category map.

4. Rust holds the desktop/utility layer. Five Rust repos in the top-100 (claw-code, cc-switch, rustdesk, rust-lang, tauri) plus herdr (agent multiplexer) trending. cc-switch specifically = Rust + Tauri for a cross-platform agent-provider desktop app, which is the exact shape of our tooling preference. See rust-agent-ecosystem and moltis-rust-agent.

5. Privacy/self-hosting is structurally in the top-100, not a niche. awesome-selfhosted (#11), open-webui (#56), immich (#98), prompts.chat marketing itself on “self-host for complete privacy”, OpenClaw’s entire pitch being “runs on your own devices.” privacy-first is a market position, not just an ethic.

The skeptical half

Stars stopped being a quality signal. claw-code sits at #27 with 195K stars and 1611★/day — the highest velocity in the entire top-100 — and its README is a badge wall pointing at two other harnesses, describing itself as “an agent-managed museum exhibit.” Several others in the cohort are single-file repos with 12-language README translations, Discord funnels, and star badges served from their own API. Rank in this chart measures distribution mechanics (X threads, agent-ecosystem cross-promotion, install one-liners), not code you’d want to depend on. Read the repo before citing the number — the same discipline as Paperclip, where the “zero-human” framing turned out to be marketing over five genuinely good mechanics.

And the underlying data just closed. GitHub restricted the stargazers API on 6 July 2026 (announced 30 June): who-starred-when is now visible only to repo admins and collaborators. Star History can no longer draw growth curves for repos you don’t own; embedded README charts for third-party repos are broken with no workaround yet. Practical consequences for us:

• Star trajectories for other people’s repos are gone. Only point-in-time totals remain — this page is a snapshot, and it cannot be reconstructed retroactively.
• Any scraper that used star velocity as a traction proxy needs a new signal. The working substitute is activity data from GH Archive — that’s how star-history’s own coding-AI leaderboard now ranks tools (PR reviews, PRs opened, pushes) instead of stars. Jun 29 – Jul 28 2026: Copilot 7 874 PR reviews, CodeRabbit 6 575, Codex 4 372, Cursor 81 216 pushes.

Currently trending (22–28 Jul 2026)

Weekly leaders, useful as an early-warning list: mattpocock/skills, OmniRoute (MIT AI gateway, 290+ providers, token compression), worldmonitor (real-time geopolitical intel dashboard), Orca (ADE for a fleet of parallel agents), pi (unified LLM API + agent loop + TUI), ponytail, Graphify (see graphify-vs-solograph), hallmark, herdr (Rust agent multiplexer), alibaba/open-code-review (deterministic pipeline + LLM agent hybrid).

Related

skills-standard · agent-toolkit-landscape · harness-engineering-summary · writing-claude-md · project-solo-factory · hermes-agent · paperclip-agent-company · agent-benchmarks · rust-agent-ecosystem · privacy-first · graphify-vs-solograph