State of GitPop H2 2026: what 12 weeks of radar data tells us about open source
Twelve weeks of weekly digest data distilled into one report. The AI infra wave has plateaued. Rust tooling is winning the developer-experience war. The agent-framework space has 3 winners and 30 losers. All numbers are from the GitPop radar.
Executive summary. Over the 12 weekly digests from 2026-W23 through 2026-W34, the GitPop radar tracked 12 consecutive weeks of movement across the open-source ecosystem. Three things happened: (1) the LLM-infrastructure wave plateaued for the first time in 18 months; (2) Rust quietly took the developer-experience category from Go; (3) the AI-agent framework space crystallized around three winners and a long tail of derivatives. All numbers below are from GitPop's PopScore over the rolling 7-day window, taken from the public radar at the date the weekly digest was generated.
This is the first State-of-GitPop report. We will publish one each half-year.
How to read this report
Every number in this report is a PopScore (0-100) for a specific project, measured at the time the corresponding weekly digest was generated. PopScore weights commits, PRs, and issue-close rate above raw star velocity and applies a noise penalty for awesome-* lists and tutorial repos. The full formula is in the methodology page. All "Δ 7d" numbers are the project's 7-day star growth at the time of measurement.
A note on sample size: this report draws on the projects GitPop actively tracks. We track 50-100 high-signal projects at any time, plus 500+ in our long-tail dataset. We are not claiming these numbers represent the entire open-source ecosystem — they represent the projects the radar has decided are worth tracking. The selection bias is intentional: this is what an active developer is likely to evaluate in 2026 H2.
1. The LLM-infrastructure wave plateaued
For 18 months straight, every quarterly read we did started with "AI infra is still growing." That sentence is no longer true.
What the radar shows for the 12-week window 2026-W23 to 2026-W34:
- The median PopScore of LLM-adjacent projects (Ollama, llama.cpp, vLLM, litellm, transformers, and the top 10 LLM frameworks) held roughly flat — within ±5% across the quarter. The previous two quarters each saw a 25-40% median lift.
- New LLM-framework entries dropped 32% quarter-over-quarter. We tracked 23 new "yet another LLM framework" repos in Q1 2026, 19 in Q2 2026, and 13 in Q3 2026 (through 2026-W34).
- But the top 5 LLM repos (Ollama, llama.cpp, vLLM, litellm, transformers) grew 8% across the same window. The winners got stronger; the long tail died.
The interpretation. The LLM-tooling space is consolidating. The winners are clear, the long tail is dying. New entrants face a much higher bar — every "wrapper around OpenAI" idea has been tried three times in the last 12 months, and the one that ships a complete runtime (sandboxing + persistence + observability) wins, not the one that ships a thin abstraction.
The implication for builders: if you are starting an LLM project in 2026 H2, expect to compete against an incumbent with 2x your velocity and a 10x head start on integrations. The exception is agent runtimes — see §4.
The implication for technical writers, hiring managers, and conference programmers: the LLM-infrastructure talk track is no longer the differentiated signal it was in 2024-2025. The signal has moved. See §2 and §4.
2. Rust quietly won the developer-experience category
Two Rust CLI tools are now in the top 50 overall by PopScore — the first time this has happened outside of ripgrep and fd. Both are productivity tools that replaced a Go or C equivalent.
| Project | PopScore (W34) | Δ 7d stars | Replaces | What it does |
|---|---|---|---|---|
atuin |
78 | +1,240 | shell history (bash/zsh history) | Magical shell history with TUI search, cloud sync, encrypted backups |
zoxide |
71 | +980 | cd |
Smarter cd that learns which directories you use most |
broot |
64 | +520 | tree |
Interactive tree-view with filtering, gitignore-aware |
eza |
58 | +410 | ls |
Modern ls with icons, git integration, themes |
The pattern is consistent: Rust is winning in developer-facing tools where performance + single binary distribution matter. The same pattern is showing up in:
- Build tools (turbopack, rolldown, biome) — all Rust, all in the radar
- Modern Unix replacements (fd, ripgrep, bat, eza, zoxide) — all Rust
- Terminal multiplexers (zellij 0.40 release pushed it into radar) — Rust
- Editor infrastructure (helix) — Rust
- New database engines (turso/libsql, glaredb) — Rust
If you are starting a new CLI tool in 2026 and the language choice is between Go and Rust, Rust is now the better default for solo developers and small teams. The startup cost is higher (borrow checker, slower iteration on the first 1,000 lines) but the runtime cost is dramatically lower, the binary distribution story is identical, and the ecosystem is mature enough that you do not pay the cost on every dependency anymore.
What is Rust not winning at, yet? Web services. Axum is good but not dominant; actix-web and warp have plateaued. The web-framework space is still JavaScript/TypeScript + Python. See §3.
3. Frontend consolidation: 3 winners, 5% long tail
Eighteen months ago there were 8 "real" React meta-frameworks. Today there are 3 that matter for new projects, and the gap between #1 and #3 is wide enough to be a strategy choice rather than a taste choice.
| Meta-framework | PopScore (W34) | Δ 7d stars | Notes |
|---|---|---|---|
| Next.js | 86 | +2,180 | Dominant, ~70% of new React app traffic; the Vercel deployment story is the moat |
| Remix / React Router 7 | 72 | +510 | Loyal base, ~15% of new projects; the web-standards model is the differentiator |
| Astro | 78 | +870 | Taking the "content-first" share that Next.js was weak at; the islands model is the differentiator |
| SvelteKit | 68 | +640 | Svelte 5 + runes pushed it back into the conversation; the small-bundle story is the differentiator |
The rest (Blitz, Redwood, Gatsby, SolidStart, Nuxt) are below 5% combined. If you are starting a new React project in 2026, the decision is Next.js vs Remix vs Astro, and the answer is "what kind of site is it":
- Next.js — interactive app from line 1, or any site that benefits from Vercel's edge network
- Astro — content-heavy site (docs, marketing, blog) where most pages are static
- Remix — data-loading-heavy app where you want the web-standards model and are happy to deploy anywhere
This is the same three-way shape the Ruby-on-Rails vs Sinatra vs Jekyll decision had in 2014, and the same consolidation is now happening in the JavaScript meta-framework space.
4. AI agents: the new vertical, with 3 winners
A category that did not exist 12 months ago: AI agent frameworks. The radar has tracked 41 distinct agent-framework repos in 2026 H1. Three have separated from the pack; the other 38 are fighting for long-tail position.
| Framework | PopScore (W34) | Δ 7d stars | What it does that is different |
|---|---|---|---|
| LangGraph | 82 | +1,420 | Stateful graph runtime, opinionated about persistence + checkpointing |
| CrewAI | 71 | +860 | Role-based agent collaboration; "AI team" mental model |
| AutoGen | 68 | +540 | Microsoft's research-flavored take; multi-agent conversation primitives |
The interesting question: is "AI agent" a real product category, or is it the new "we put a chatbot in our SaaS"? Our read: it is real, but the long tail is going to be brutal. The frameworks that win will be the ones that ship opinionated runtimes — sandboxing, persistence, observability — not the ones that ship LLM-friendly abstractions. The runtime is the moat; the abstraction is the table stakes.
The 38 other agent frameworks (and there are at least 38) are all variations on "call an LLM in a loop with a tool list." They will merge into the three above, or be replaced by the next wave of opinionated runtimes, in the next 6-12 months.
For decision-makers: pick one of the three above. Building on the long tail is a 12-month bet you do not want to make in 2026 H2.
5. The data layer: Postgres won
The data-layer story is the most settled of any category in this report. Postgres + a vector extension + Redis is the new default stack. The radar numbers for the data-layer category over the 12-week window:
| Engine | PopScore trend (W23 → W34) | Direction |
|---|---|---|
| Postgres (and adjacent: Supabase, Neon, pgvector, Electric SQL) | +12% | ↑ winning |
| SQLite (and adjacent: libsql/turso) | +8% | ↑ growing (edge + mobile) |
| Redis | +3% | → steady |
| MongoDB | -22% | ↓ losing |
| MySQL | -8% | → slow decline |
| ClickHouse | +18% | ↑ winning (analytics-specific) |
| DuckDB | +24% | ↑ winning (in-process analytics) |
The interpretation. "Polyglot persistence" is dead. Teams are consolidating around Postgres for OLTP, Redis for cache/queue, and either ClickHouse or DuckDB for analytics. Pick this stack for a new project in 2026 H2 and you will not regret it for 2-3 years.
The Postgres losers are the SQL-first-databases that have not yet shipped a vector story (MySQL) and the document stores whose flexibility no longer outweighs their lack of relational integrity (MongoDB). The new winners are the OLAP-specific engines (ClickHouse, DuckDB) — they are not displacing Postgres, they are complementing it.
6. What this means for selection decisions
This is the practical upshot. If you are picking tools in 2026 H2, the categories above tell you:
- For an LLM-application backend: you are choosing between LangGraph, CrewAI, AutoGen. Do not pick a long-tail framework. (compare matrix has 2 of the 3 ranked; the third is in the radar.)
- For a CLI tool: Rust first, Go second, anything else a distant third.
- For a new web app: Next.js (if interactive) or Astro (if content-heavy). Remix if you have a strong web-standards preference.
- For data storage: Postgres. Add Redis for cache. Add DuckDB or ClickHouse if you need analytics.
- For an "AI agent" project: pick a runtime, not an abstraction. The opinionated runtimes will eat the thin abstractions by 2027 H1.
The full compare matrix has 30 high-volume "X vs Y" decision pages for the categories above. The projects directory has the full tracked set. The methodology page has the math.
7. What this report is not
- Not a leaderboard. PopScore measures velocity, not quality. A project at PopScore 70 is not "better" than a project at PopScore 50 — it is gaining momentum faster right now. Some of the best projects in the world have low PopScore because they are not in their growth phase.
- Not a market forecast. The "LLM wave plateaued" claim is a description of what our radar saw in 12 weeks, not a prediction about 2027. The LLM wave could re-accelerate; the agent wave could collapse. Use this as one data point among many.
- Not a substitute for due diligence. Every project linked in this report is one click away from its full PopScore breakdown, AI snapshot, star history, and external mentions. The point of the radar is that no claim is hidden.
8. Methodology + reproducibility
All numbers in this report are from the GitPop radar, drawn from the weekly digests for 2026-W23 through 2026-W34. The weekly digest itself is auto-generated from the live data and re-validated every 6 hours. The full PopScore formula and the noise filter are public.
If you want to challenge any number in this report, the path is:
- Open the linked project page
- Read the methodology page for the math
- Read the blog post on noise filtering
- Compare against the weekly digests for the date in question
Every input is auditable; no black boxes. That is the whole point.
Data source: GitPop PopScore, 2026-W23 to 2026-W34, 12 weekly digests, ~80 tracked projects across 8 categories. Generated: 2026-08-26. Re-run: The full dataset is live at /projects. The radar updates every 6 hours.