165,000 stars: open source is quietly eating the AI coding agent market

Hey — Guy here. Two weeks ago Microsoft killed Claude Code internally. Last week Google bet all of I/O on agents. This week the most interesting thing in AI coding didn't come from any of the giants — it came from a GitHub repo with 165,000 stars and no marketing budget. Open source is quietly eating the terminal while everyone watches the trillion-dollar players trade blows. ## 🧠 This Week in AI Coding - **OpenCode crossed ~165,000 GitHub stars** — and is now openly called "the de facto open-source Claude Code." Provider-agnostic, runs local models via Ollama, polished terminal UI, 7.5M+ monthly devs. Built by a tiny team (the terminal.shop folks), not a lab. The lesson: you don't need a frontier model to win the *harness* layer. - **OpenWork hit 15.5K stars as an open-source Claude Cowork alternative** — a local-first desktop app built *on top of* OpenCode that lets teams ship repeatable agentic workflows. The proprietary-vs-open race that defined browsers in the 2000s is now happening for agents, on fast-forward. - **Anthropic's "Cowork" keeps expanding** — 20+ MCP connectors and 12 practice-area plugins (legal vertical first) shipped open source. Anthropic is betting the *connectors* are the moat, not the app. Watch this pattern: give away the shell, monetize the model + the integrations. - **Cursor shipped Composer 2.5** — its in-house long-horizon model that now matches Opus 4.7 on benchmarks. Cursor no longer just *wraps* frontier models; it's building its own. That's a real strategic shift for a company that started as a VS Code fork. - **The market share story is now official:** Copilot fell from 67% → 51% among pro devs, Cursor debuted at 18%, Claude Code hit 10% on its first survey appearance. The Microsoft monopoly is genuinely over — three tools within arm's reach of each other. - **GitHub's billing change lands June 1** — Copilot moves to usage-based pricing and paused new Pro/Pro+ sign-ups. The flat-rate era I wrote about a few weeks ago is officially closing across the board. - **Google's Gemini 3.1 Ultra is shipping with a 2M-token context window** (native text/image/audio/video) plus Gemini Spark, a general agent in the Gemini app. The context arms race didn't stop — it just moved up an order of magnitude. ## 🚀 What We Shipped at VeriduxLabs Building in public, so here's the honest version: - **Opened a permanent mobile-native lane.** Turns out we already had *five* mobile codebases sitting in the portfolio. First three to ship: **Zmani** (a Hebrew halachic-times calendar — Expo + React Native, Play Store assets ready, launch imminent), then **Plate Penalty** (food-photo → calorie estimate via Gemini Flash), then a SPARit mobile companion. Sometimes the move isn't building more — it's shipping what's already 80% done. - **VCA makeover handed off to a coding agent.** Wrote a full self-contained audit-and-rebuild brief (12 areas, 3-phase ship framework) for Cursor to take Vibe Coding Academy end-to-end. New `CURSOR-TASK-*` handoff convention — three different AI agents now coordinate through files in one repo. - **First-dollar focus locked.** Migrating VCA payments from dormant Paddle → live Lemon Squeezy and activating the course gate. One lever this week; everything else runs on autopilot. - **YouTube:** backlog fully cleared on @endofcoding — every rendered video is now live. Next batch in production. ## 🛠 Tool of the Week: OpenCode If you've been Claude-Code-curious but allergic to vendor lock-in, this is your week. **OpenCode** is a free, MIT-licensed, terminal-native coding agent that talks to 75+ model providers — Anthropic, OpenAI, Google, *or* a fully local model through Ollama. Same agentic loop (read files, run commands, edit code, verify), zero subscription, and you own the whole stack. Mini-review: the TUI is genuinely good, switching models mid-task is one keystroke, and running it against a local model on a beefy laptop means you can code agentically on a plane with no API bill. It's not quite as polished as Claude Code on the hardest long-horizon tasks yet — but for 90% of day-to-day work, the gap is smaller than the price difference suggests. Try it: `opencode.ai`. ## 🔗 Quick Links - [OpenCode](https://opencode.ai/) — the open-source agent that hit 165K stars - [OpenWork on GitHub](https://github.com/different-ai/openwork) — open-source Claude Cowork alternative - [The New Stack: Cursor, Claude Code & Codex are merging into one stack nobody planned](https://thenewstack.io/ai-coding-tool-stack/) - [Awesome CLI Coding Agents](https://github.com/bradAGI/awesome-cli-coding-agents) — the full terminal-agent landscape, curated --- That's the week. The takeaway I keep coming back to: in 2026, the *model* is increasingly a commodity and the *harness* is where the real fight is — and open source just proved it can win the harness without a lab behind it. Worth watching where your tooling loyalty actually sits. See you next Sunday, **Guy** --- 📺 [YouTube @endofcoding](https://youtube.com/@endofcoding) · 📘 [vibecodingebook.com](https://vibecodingebook.com) · 💼 [llmhire.com](https://llmhire.com) · 🛡 [cyberos.dev](https://cyberos.dev) · 🌐 [endofcoding.com](https://endofcoding.com) *Built in public with a team of AI agents. Reply anytime — I read every one.*