LearnWithLeon

@jg_noncelogic @roveapi taprun.dev

Shipped: @roveapi — a Playwright API built specifically for AI agents. Most browser APIs return screenshots. Screenshots cost ~114K tokens per page. Rove returns accessibility trees instead. ~26K tokens. Same data. MCP-native so Claude, Cursor, and VS Code agents can use it without writing integration code. Free tier live now → roveapi.com

@gregpr07 taprun.dev

Introducing: Browser Harness. A self-healing harness that can complete virtually any browser task. ♞ We got tired of browser frameworks restricting the LLM. So we removed the framework. > Self-healing — edits helpers. py on the fly > Direct CDP — one websocket to Chrome > No framework, no rails, complete freedom > Drop-in for Claude Code and Codex I challenge anyone to find a task that DOESN'T work. I couldn't yet.🔥 100% open source ↓

@an_hyoung 100% this. Local-first dev tools are underrated — no reason your code should leave your machine just to format JSON. Nice work building all 8 tools client-side.

Every time you paste code into an online JSON formatter, it gets sent to their server. I built CodeNeat — 8 developer tools that run 100% in your browser. Your code never leaves your machine. Free and open source. codeneat.dev

200+ skills: github.com/LeonTing1010/t…

@jancurn @AgenticAIFdn Without that, you're handing strangers a remote browser.

On my way back from the MCP Dev Summit hosted by @AgenticAIFdn in NYC. Looks like MCP is not so dead.

@Shruti_0810 The key insight: the runtime is swappable, but the program (.tap.js) stays the same. Write once, run on any backend.

Best GitHub repos for Claude Code (Top 10): Claude Code (Official) github.com/anthropics/cla… Run Claude in your terminal → build + edit projects end-to-end. Claude Skills github.com/anthropics/ski… Turn repeat tasks into reusable workflows. Claude Code GitHub Action github.com/anthropics/cla… PR reviews, bug fixes, auto-shipping code. Awesome Claude Code github.com/hesreallyhim/a… Real setups. Real agents. No fluff. Awesome Claude Skills github.com/travisvn/aweso… Plug-and-play skills for debugging + scaffolding. Composio Claude Skills github.com/ComposioHQ/awe… Production-ready workflows + integrations. LangGraph github.com/langchain-ai/l… Multi-step agents with memory. Microsoft AutoGen github.com/microsoft/auto… AI dev teams that plan → code → review. CrewAI github.com/crewAIInc/crew… Autonomous agents collaborating like real engineers. Continue github.com/continuedev/co… AI copilot inside VS Code / JetBrains.

@jason_haugh @mariotomich We built Tap to fix this — AI forges a deterministic program once, then it replays forever without AI. The trace is diffable, so when a site changes, you see exactly what broke.

@mariotomich I run my entire AI agent infrastructure off a Mac Mini at home. 24/7, six specialized agents, cron jobs, browser automation, the whole stack. It paid for itself in the first month. People spending more on coffee have no room to judge.

I really don't get this constant bashing of people who buy DGX Sparks, Mac Studios, or GPUs to build local AI setups. People spend the equivalent of a DGX Spark on junk food and Starbucks every year. But investing in learning and building with the most disruptive technology of our time is unreasonable?

@sidelined_cap tap doctor + health contracts make it automatic: if min_rows drops to 0, you know the site changed. Normal engineers can own this, no prompt wizardry required.

@LearnWithLeon Love the POSIX framing. The practical unlock is deterministic replay: same primitive trace, diffable failures. That makes UI agents debuggable by normal engineering teams, not just prompt wizards.

@m13v_ @mercer70638 @tom_doerr Self-healing selectors are a step in the right direction but they still burn AI tokens on every run.A more radical approach: AI writes a complete deterministic script once, with health contracts (expected row count, non-empty fields). When the contract breaks, THEN re-invoke AI.

@mercer70638 @tom_doerr DOM structure changes breaking automation is the #1 pain point. self-healing selectors help a lot, they use multiple strategies (text, role, position, test-id) and auto-adapt when one breaks. been building this into an open source Playwright test generator

AI browser automation for Claude Code github.com/DebugBase/glan…

@sidelined_cap Exactly — the trace is the log. When something breaks, you diff the trace, not the prompt. `tap doctor` + health contracts make it automatic: if min_rows drops to 0, you know the site changed. Normal engineers can own this, no prompt wizardry required.