10 Trending AI Dev Tools, Week of April 28 2026

A real April, not a recap of last quarter

The last 30 days were the loudest month of 2026 so far. Anthropic shipped Opus 4.7. OpenAI shipped GPT-5.5 and rebuilt the Codex CLI. Google shipped Gemini 3 Pro and Antigravity. The Claude Code skills ecosystem went from "neat side trick" to a mainstream extension model with thousands of public skills. And on the open-source side, two repos (Andrej-karpathy-skills and Hermes Agent) absorbed most of the GitHub Trending oxygen.

Here are the ten tools, models, and repos worth your attention this week, with the commands and links you actually need.

1. Claude Opus 4.7

Released April 16, 2026. On Anthropic's 93-task internal coding benchmark, Opus 4.7 lifted resolution by ~13% over Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve. Pricing stays at $15 / $75 per million tokens. Available in Claude apps, the API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Pin it in Claude Code:

claude --model claude-opus-4-7

Or set it as the default in ~/.claude/settings.json:

{
  "model": "claude-opus-4-7"
}

2. OpenAI GPT-5.5 (and the new Codex CLI)

GPT-5.5 hit the API on April 24, 2026. OpenAI's positioning is "smarter and more token-efficient than GPT-5.4" - and inside Codex specifically, it produces better diffs with fewer tokens.

The bigger story is Codex CLI itself. Recent Codex changelog entries added Unix socket transport for the app-server, sticky environments, remote plugin install, automatic reviewer agents for risky approvals, and codex exec --json reasoning-token reporting. Try the new browser hand-off:

codex
> use the browser to reproduce the layout bug on localhost:3000

If you have not picked sides yet, our Claude Code vs Codex App breakdown is still mostly accurate - just bump every Codex bullet up a notch.

3. Gemini 3 Pro and Google Antigravity

Google released gemini-3-pro-preview on April 22, 2026 with strong agentic and coding behavior. The Gemini 3 developer guide covers the new tool-use loop. The more interesting piece is Google Antigravity, an agentic IDE-adjacent platform where Gemini 3 plans and executes end-to-end tasks while self-validating.

Gemini 3 Flash arrived a few days later and is already wired into the Gemini CLI:

gemini -p "summarize the diff in HEAD~1..HEAD"

4. Claude Haiku 4.5

If you missed it earlier this quarter, Haiku 4.5 is still the price/performance pick: roughly Sonnet 4-tier coding at one third the cost and twice the speed. Pricing is $1 / $5 per million tokens. Worth slotting in for cheap parallel sub-agents while reserving Opus 4.7 for the planner.

5. Andrej-karpathy-skills

The breakout repo of late April. As covered in GitHub Trending Weekly, it added roughly 44k stars in a week and triggered the wider Claude Code skills ecosystem boom. The skills are pedagogical and surprisingly good as starter material. Drop them into ~/.claude/skills/ and they show up automatically.

gh repo clone andrej-karpathy/skills ~/.claude/skills/karpathy
ls ~/.claude/skills/karpathy

6. The Claude Code skills marketplace

claudemarketplaces.com and the open-source claude-code-plugins-plus-skills (423 plugins, 2,849 skills, 177 agents) became real this month. A plugin bundles skills, MCP servers, slash commands, and sub-agents into one installable unit, which is the right granularity. If you have not started writing your own, our Claude Code skills guide is the fastest way in.

7. Hermes Agent

Hermes Agent hit 65k stars and added 32k of those in a single week earlier in April. It is one of the more enduring agent frameworks on Trending - good defaults, simple Python API, and plays nicely with MCP servers. Worth a look if LangGraph feels heavy.

8. EvoMap and self-evolving agents

EvoMap and GenericAgent pushed self-evolving agents into mainstream attention. EvoMap's "Genome Evolution Protocol" lets agents mutate and select prompts, tools, and policies over time. GenericAgent grows a skill tree from a 3.3k-line seed and claims roughly 6x lower token use than static frameworks. Both are early, both are interesting, neither is production-ready yet.

9. MarkitDown

Microsoft's MarkitDown is still adding ~7k stars per week. It converts PDF, Word, Excel, PowerPoint, and HTML to clean Markdown so any LLM pipeline can ingest them. It's the boring, useful kind of tool you should already have installed:

pipx install markitdown
markitdown report.pdf > report.md

This is the kind of unglamorous glue that quietly shows up in every serious agent stack.

10. The visual builders, still huge

Three of the top five AI repos on GitHub are visual workflow builders: Langflow at 146k stars, Dify at 136k, and Flowise at 51k. n8n crossed 180k stars. They are not what you'd reach for to build a hard agent product, but they remain the fastest way to wire a prototype that calls real APIs and real models.

What to actually do this week

If you have one hour, do three things:

1. Switch your default Claude Code model to Opus 4.7 and your sub-agent model to Haiku 4.5. This is the highest-value change of the month for most repos.
2. Update Codex CLI (npm i -g @openai/codex@latest) and try codex exec --json against a small repo to see GPT-5.5's reasoning-token output.
3. Drop one community skill into ~/.claude/skills/ and watch how Claude picks it up automatically. Once you do this once, you write your own skills the next day.

The rest of the list is worth keeping tabs on, but those three changes alone will move how your agents behave on Monday.

For deeper picks see Best MCP servers 2026 and Best CLI tools for AI development 2026.