Six clear themes emerged from Product Hunt and GitHub Trending this week. Below, the tools clustered by what they actually have in common — not just their category tags.

The week's most-upvoted Product Hunt launches shared one trait: they execute on your behalf inside the tools you already use, rather than requiring you to open a new tab and type a prompt.

Viktor 1 is the most fully realized version of this idea. It lives inside Slack, connects to 3,000+ tools across your stack, and does work proactively — it watches how your team operates, spots repeated problems, and proposes automations before anyone asks. It runs for weeks without losing context. Pricing is not listed publicly; teams can request access at getviktor.com. The differentiator against Zapier or Make: those tools handle clean, predictable triggers; Viktor handles judgment calls.

Yansu 2 takes a similar observation-first stance but focuses on pattern recognition across your existing files, messages, and workflows. It spots what you do repeatedly and turns those patterns into apps and automations — without any process mapping or blank-canvas configuration. Free options are available; full pricing is on yansu.app. The key difference from generic no-code builders: you never start from zero.

Coworker AI 3 comes at the same problem from a cost angle — context-aware model routing that automatically picks the cheapest model capable of handling each specific task. If Viktor and Yansu are about eliminating manual work, Coworker AI is about eliminating overspending on LLM compute.

Try it: Give Viktor your most repetitive Slack task. Give Yansu a look at an ops folder you haven't automated yet. Both have low setup overhead.

Persistent memory — how an AI carries knowledge of you across tools and sessions — broke out as a standalone category this week. Two very different approaches landed simultaneously.

Unabyss 4 attacks the problem at the API layer. It's MCP-native: connect your daily apps once, and Unabyss extracts, structures, and continuously updates a personal context vault. Any MCP-compatible AI tool can then read that context — with granular controls over what each tool can actually see. Free at launch. The founder framed the problem sharply: "ChatGPT memory doesn't follow you to Claude. Claude Projects don't talk to Cursor." Unabyss is the routing layer between them.

On GitHub, plastic-labs/honcho 5 is a memory library aimed at developers building stateful agents — 708 stars gained this week on top of its 4,400 total. Where Unabyss serves end users managing their own context, Honcho is the infrastructure devs drop into their agent pipelines.

Anthropic's knowledge-work-plugins 6 (4,718 new stars, 17,345 total) takes an open-source approach: a catalog of plugins for Claude Cowork aimed at knowledge workers. Think of it as a curated app store for Claude's context capabilities, maintained by Anthropic itself.

Try it: If you use 3+ AI tools daily, Unabyss is worth the 10-minute setup. If you're building agents, Honcho deserves a read before you roll your own memory layer.

Autonomous agents running in production raised a question this week: who's watching them? Two projects emerged directly aimed at that gap.

microsoft/agent-governance-toolkit 7 (1,314 new stars) covers policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous agents — and claims coverage of all 10 items in the OWASP Agentic Top 10. Open source, MIT-adjacent license.

mukul975/Anthropic-Cybersecurity-Skills 8 (4,170 new stars, 11,046 total) is a more unusual artifact: 754 structured cybersecurity skills for AI agents, mapped to five frameworks including MITRE ATT&CK, NIST CSF 2.0, and NIST AI RMF. It works as a drop-in skill set for Claude Code, GitHub Copilot, Codex CLI, Cursor, and 20+ other platforms. The parallel growth of both repos suggests practitioners are actively assembling security stacks for agent deployments, not just reading about the problem.

Try it: If you're deploying agents in a regulated environment, microsoft/agent-governance-toolkit is a faster starting point than writing policy from scratch.

Voice agents came up in multiple places this week — and the consistent theme was "self-host or bring your own stack."

Dograh 9 (997 new stars) is a direct open-source alternative to Vapi and Retell. It's BSD 2-Clause licensed, self-hostable via a single Docker command, and BYOK across the full speech-to-speech or LLM/STT/TTS stack. Built-in telephony (Twilio, Vonage, others), a visual workflow builder, and native MCP support. Self-hosted: free. Cloud: usage-based. The YC alumni team behind it explicitly targets teams locked into Vapi's pricing or data residency constraints.

Parrot Speech-to-text API 10 landed on Product Hunt as a purpose-built STT layer for production voice agents — prioritizing accuracy and latency over general-purpose transcription. Pricing is on their site; the differentiator is the focus on agent-grade reliability rather than personal transcription.

Try it: If you're evaluating Vapi, spin up Dograh locally first. The Docker path is under 2 minutes.

rohitg00/ai-engineering-from-scratch 11 was the biggest GitHub mover of the week: 12,787 new stars, 22,702 total. The tagline is "Learn it. Build it. Ship it for others." — a practical engineering curriculum for building AI systems from first principles, not a paper survey or a list of links.

Powabase 12 sits at the other end of the spectrum: a ready-to-use platform for building AI apps with Postgres, RAG pipelines, and agents — no infrastructure assembly required. Positioned for developers who want production-ready agent infrastructure without configuring vector stores by hand.

Supaboard 3.0 13 addresses a narrower but common problem: AI data analysts that understand your specific business context, not just SQL syntax. The v3.0 release suggests they've iterated past the "AI writes queries" phase into something that models business logic.

Try it: ai-engineering-from-scratch is worth bookmarking even if you're not starting from zero — the repo structure suggests it works well as a reference.

The video generation space moved toward agents this week — systems that don't just generate a clip but handle the whole production process.

HKUDS/ViMax 14 (1,940 new stars) frames itself as "Director, Screenwriter, Producer, and Video Generator All-in-One" — an agentic pipeline for video generation where the AI handles narrative structure, not just individual frames.

AVTR-1 15 is Avaturn's open-source real-time model for uncanny AI avatars — the first open-weights model in this category. It opens up avatar generation to anyone running their own stack, rather than requiring an API call to a commercial service.

CloakHQ/CloakBrowser 16 is adjacent: a Chromium fork that passes all standard bot detection tests (30/30), built as a drop-in Playwright replacement. 4,427 new stars. Not purely an AI tool, but heavily used in AI agent pipelines that need to navigate the web without being blocked.

Try it: If you're building browser-based agents, CloakBrowser is worth testing against your current automation stack.

Two repos that don't fit cleanly into the above themes are worth flagging. Imbad0202/academic-research-skills 17 (7,385 new stars) structures Claude Code's research workflow across phases: research → write → review → revise → finalize. And aiming-lab/AutoResearchClaw 18 (451 new stars) goes further — "chat an idea, get a paper" — a self-evolving research system that takes an idea as input and produces a full academic paper.

Both repos reflect a pattern: the tooling for getting AI to do sustained, structured knowledge work is advancing faster than the tooling for getting it to do any single task well.

Sources this week: Product Hunt AI leaderboard, week of May 25 · GitHub Trending Python, this week