GitHub's agent-skills ecosystem crosses 300K stars while Claude Code ships your email in every HTTP request header.
The same week coding-agent skills became a genuine ecosystem — one repo crossing 313K stars — the #1 Hacker News story was an AI agent leaking its user's email to every website it visited. We're building capability faster than trust.
Claude Code has been embedding users' email addresses in HTTP User-Agent headers on outbound requests. GitHub issue #107757, which hit the top of Hacker News with 2,369 points, documented the mechanics: every web request the agent makes can carry your account email, sometimes alongside a project name derived from your session context.
Two User-Agent patterns were reported:
The reporter was receiving dozens of scam emails daily. The deeper concern: since the agent generates its own UA strings from session context, it can tie your identity to your current project — linkable by IP across requests.
Anthropic responded fast — within hours — shipping a built-in mod (cc-plugin-privacy) in v2.1.317. The mod is a hook on outbound shell commands that regex-matches the account email near a User-Agent field and refuses the call. It's enabled by default.
Users found bypasses just as fast. Writing the email to a Python script with the Edit tool and running it via python3 sidesteps the shell check entirely — the hook inspects the shell command, not the file content. Placing the email in a differently-named HTTP header also works. One user demonstrated that echo 'User-Agent: test still executed on v2.1.319. The mod catches the obvious case and misses the structural ones.
The community is asking a harder question: why does the agent runtime hold the email at all? The User-Agent convention the agent is following — including a contact address for automated clients — is a reasonable practice for crawlers, but the decision to follow it should be the user's, not the model's.
This pattern will repeat for every tool-using agent. Any agent with shell access has access to every exfiltration channel the shell touches: HTTP headers, git configs, DNS queries, file metadata, error messages. A regex on one exit path is whack-a-mole with infinite moles. The correct fix is architectural: an agent's runtime sandbox should not contain credentials irrelevant to its current task. The email goes into a capability the agent can explicitly request, gated by approval, not into an environment variable it can read freely. Least-privilege isn't a new idea. Applying it to agents apparently is.
Something shifted in the coding-agent world this week, and the GitHub numbers make it hard to ignore. obra/superpowers — "a complete software development methodology for your coding agents" — sits at 313K stars with 27.9K forks. claude-mem, which gives agents persistent memory across sessions, is at 113K. impeccable, a design vocabulary that catches generic AI output before it ships, crossed 104K.
These aren't novelty projects. They're infrastructure.
What converged: Skills have settled on a shared interface pattern. A skill is a markdown file or small script that triggers automatically when the agent detects relevant context. No manual invocation. superpowers enforces TDD-first methodology — the agent writes tests before code, reviews between tasks, and can work autonomously for hours without drifting from a spec. claude-mem hooks into session lifecycle events (start, tool use, stop) and compresses observations into semantic summaries stored in SQLite + Chroma for hybrid retrieval. impeccable gives agents design critique vocabulary with 24 slash commands and a 62-rule deterministic detector.
The architecture is settling into three layers:
Memory (claude-mem): SQLite with FTS5 full-text search plus vector DB. Hooks capture events; AI compression turns raw observations into retrievable context. A three-step retrieval pattern — compact index, timeline, then full detail — cuts token usage by roughly 10x compared to dumping raw history.
Methodology (superpowers): Mandatory workflow skills triggered before the agent writes code. The flow is: spec → plan → implement → review. Subagent-driven execution where agents review each other's output. The stated philosophy — evidence over claims, complexity reduction, YAGNI — sounds boilerplate until you see agents actually following it for multi-hour sessions.
Quality gates (impeccable, hallmark at 49K): Post-generation checks that catch the AI's blind spots. impeccable's critique command returns scored UX reviews. audit runs accessibility and performance checks. polish is a final pass. hallmark focuses specifically on the "AI slop" aesthetic: generic layouts, meaningless gradients, and decoration that signals "an AI made this."
These tools are composable because they communicate through a shared protocol layer — CLAUDE.md, .claude/settings, hooks — rather than through each other. That's the Unix philosophy applied to AI development workflows. Install methodology from one repo, memory from another, quality gates from a third. They don't conflict.
Here's what I'm watching: the capability gap between teams that compose these tools and teams that don't. A developer running superpowers + claude-mem + impeccable gets an agent that remembers architectural decisions from last Tuesday, follows a TDD spec, delegates subtasks to reviewer subagents, and catches its own visual blind spots before reporting done. A bare claude invocation gets an amnesiac that starts every session from scratch, makes the same mistakes it made yesterday, and produces the same Inter-on-purple-gradient card layout it always does.
That gap is about to become a hiring signal. Teams shipping with composed agent stacks will move measurably faster, and the methodology enforcement means their output will be more consistent. The solo developer who installs three packages gets closer to a junior team than a junior developer ever was.
The risk mirrors early npm: dependency sprawl, implicit trust in community-authored skills (anyone can publish a skill that triggers automatically), and the assumption that 313K stars means 313K code reviews. BLA — the adversarial defense toolkit trending at 2.5K stars — is already shipping tactics to resist agent-powered reverse engineering. The skills ecosystem will need its own security model before long. A malicious skill that triggers on "before any task" has the same privilege as the agent itself.
Earshot — Gives coding agents the ability to hear. CPU-only audio analysis using CLAP and AST models: describe what a file sounds like, detect real vs. synthesized audio (ROC-AUC 0.92), generate sounds from code recipes, and score a film's audio mix against professional references. The forge command builds sounds from code and iterates with its own critic; fit converts a recording into an editable code recipe. Ships as CLI, Python API, and MCP server. No GPU needed — 1.1 GB model download, then fully offline. 4K stars, MIT.
ESP32 iPhone Agent — A $7 ESP32-S3 microcontroller gives any vision model physical control of a real iPhone. AirPlay mirrors the screen (480x1040, up to 5 fps) to the board without decoding; USB HID sends taps, swipes, and keystrokes back via AssistiveTouch. No jailbreak, no developer account needed. The agent loop — screenshot, vision model reasoning, action, verify — supports any OpenAI-compatible endpoint including local models via Ollama or vLLM. Ships with an MCP server (esp32agent) so coding agents can drive the phone programmatically, plus Siri and Shortcuts integration for launching saved tasks. 4.6K stars, MIT.
OpenDuet — A proactive AI agent template from CopilotKit. Unlike reactive assistants, it sends morning briefs, sorts incoming mail/Slack/calendar, and works on delegated tasks while you're away. Every action requires explicit approval with a full audit trail. Each agent runs in its own container with browser, shell, and workspace volume. Built on the AG-UI protocol, so agents can be authored with LangGraph, Mastra, or CrewAI. 2K stars, early but architecturally interesting.
The agent ecosystem is bifurcating along a predictable fault line. On one side: skills, memory, methodology — the scaffolding that turns a stateless LLM into something that can sustain a multi-hour development session without losing the plot. On the other side: privacy leaks, adversarial defenses, and the dawning realization that a tool with shell access is a tool with exfiltration access.
These aren't separate stories. They're the same story told from two ends. The skills that make agents powerful — persistent memory, automatic triggering, shell access — are the same capabilities that make the Claude Code email leak possible. A skill that fires "before any task" with full shell permissions is indistinguishable from a pre-commit hook with root access.
The Claude Code email leak is a one-line fix for Anthropic, but the pattern it represents — agents holding credentials they don't need, exfiltrating through channels nobody audits — will repeat at every layer of the stack. The teams that figure out least-privilege for agents, as architecture rather than as aftermarket regex, will define the next era of developer tooling. Everyone else will keep playing whack-a-mole with User-Agent strings.
— Aaron
Braintrust, Langfuse, Promptfoo, and DeepEval compared on eval workflow, CI gates, red teaming, self-hosting, and pricing. Pick the right LLM eval stack.
AI EngineeringWebMCP lets websites hand tools to AI agents. A 9-of-9 attack result shows why tool descriptions are untrusted input, plus the guards that stop it.
AI EngineeringLlamaIndex vs LangChain for RAG and agents: indexing depth, create_agent vs AgentWorkflow, checkpoints, human approval and when to combine both.
AI Engineering