New AI Tools Weekly — Issue #3: Six Themes from June 8–14, 2026

The sharpest pattern in this week's data: the AI tooling layer is thickening below the model. Context compression, agent scaffolding, voice synthesis primitives, open research workspaces — these aren't the "what model should I pick" conversation. They're the plumbing conversation. The week's biggest GitHub movers all sit in that plumbing layer, while Product Hunt's weekly chart tilts heavily toward agent infrastructure rather than end-user apps.

Fifteen tools across six themes. What's new, what costs what, and where each one fits.

Two of this week's top three GitHub movers tackle the same hidden tax on agentic workflows: the tokens you pay to shovel context to the model, most of which are waste.

headroom (chopratejas/headroom, +14,272 stars this week, 17.4k total) is a context compression layer that sits between your agent and the LLM provider. It routes content through type-specific compressors — SmartCrusher for JSON, CodeCompressor for ASTs, Kompress-base (a local HuggingFace model) for plain text — and reports 60–95% token reduction on real agent workloads. Benchmarks from the repo show a 100-result code search dropping from 17,765 tokens to 1,408 (92% reduction), and an SRE troubleshooting session from 65,694 to 5,118 tokens — with GSM8K accuracy held at 0.87 before and after compression.1

It ships as a Python library (pip install "headroom-ai[all]"), an npm package, a Docker image, and an MCP server with three tools (headroom_compress, headroom_retrieve, headroom_stats). A proxy mode (headroom proxy --port 8787) wraps any OpenAI-compatible client with no code changes. License is Apache 2.0, free. The only real caveat: the local ML model (Kompress-base) adds a setup step and assumes you have Python 3.10+ and enough RAM to run a small HuggingFace model locally.

microsoft/markitdown (+13,359 stars, 147.7k total) is not new — it's been building steam for months — but the week's star acceleration suggests it's crossing into mainstream agent pipelines as a document-to-context preprocessing step. It converts PDFs, Office docs, images, audio, HTML, CSV, JSON, XML, and ZIP archives into Markdown, which then feeds cleanly into LLM context. Free, MIT license, installable via pip install markitdown. The differentiation is breadth: where PyMuPDF or pdfplumber handles PDFs only, markitdown handles the whole office document zoo in one package.2

Together these two define a workflow many agentic teams are quietly standardizing on: markitdown converts raw files to Markdown, headroom compresses that Markdown before it reaches the LLM.

NousResearch/hermes-agent (+11,427 stars this week, 187k total) has been building since Issue #1's context/memory theme. This week the community attention spiked hard, likely driven by the companion nesquena/hermes-webui (+4,281 stars, 13.9k total) landing and making the agent accessible to non-CLI users.

The core differentiator from Claude Code or Cursor isn't code editing — Hermes is built for persistent autonomous operation. It creates skills from task experience, retrieves them across sessions using SQLite FTS5 full-text search, builds a user model via Honcho, and runs unattended cron jobs described in natural language. Deploy targets include 5-dollar VPS instances, GPU clusters, and serverless (sleeps when idle). The agent connects to Telegram, Discord, Slack, WhatsApp, and Signal via a message gateway, so tasks you schedule on a VPS can push results to wherever you actually check notifications.3

There's no model lock-in: Hermes routes through Nous Portal, OpenRouter (200+ models), NovitaAI, NVIDIA NIM, Kimi, MiniMax, or a custom endpoint. The MIT license means commercial use is unrestricted. The optional Nous Portal subscription adds cloud browser, image generation, and TTS; self-hosting with your own API keys is fully supported at no extra cost.

hermes-webui by nesquena gives non-terminal users a browser-based Hermes interface. At 13.9k stars, the web layer is tracking toward the core repo at a pace that suggests the audience for Hermes is broader than early adopters assumed.

OpenBMB/VoxCPM (+4,298 stars this week, 27.6k total) ships VoxCPM2, a 2B-parameter TTS model that skips discrete tokenization entirely. Instead of converting speech to a token sequence, it runs a diffusion autoregressive pipeline (LocEnc → TSLM → RALM → LocDiT) operating directly in the latent space of its AudioVAE V2. The practical result: 48kHz studio-quality output from a 16kHz reference clip, no external upsampler needed.4

The benchmark table in the repo shows VoxCPM2 outperforming CosyVoice3 (0.5B and 1.5B), F5-TTS, and IndexTTS2 on the Seed-TTS-eval English WER metric, while matching or beating MOSS-TTS on most tracks. On the InstructTTSEval Voice Design benchmark (designing a voice from a text description alone), VoxCPM2 scores 84.2 APS and 83.2 DSD in English — best among open models in that eval.

Key capabilities: 30-language support, voice design from text description only (no reference audio), controllable cloning with style guidance, and real-time streaming at RTF ~0.3 on an RTX 4090 or ~0.13 with vLLM-Omni. Apache 2.0 licensed, free for commercial use. The main barrier is model size: 2B parameters means a GPU with at least 8–10GB VRAM for practical inference.

Open-LLM-VTuber/Open-LLM-VTuber (+2,388 stars, 10.4k total) shows up as a related thread: hands-free voice conversation with any LLM, voice interruption support, and a Live2D talking avatar running locally across platforms. It's the integration layer on top of models like VoxCPM2 — useful for developers building interactive voice characters or local assistants with visual presence. MIT licensed.

Three tools this week are building infrastructure for agents to do things, rather than building agent reasoning itself.

Browse.sh (Browserbase, on Product Hunt Week 24) pitches "muscle memory for web automation" — a persistent browser context layer that lets agents accumulate learned interaction patterns for specific sites, rather than re-discovering navigation paths each session. Categories: API, Developer Tools, AI. Pricing wasn't shown on the leaderboard listing; the project page hit a Cloudflare block during data collection.5

Panniantong/Agent-Reach (+2,289 stars, 23.4k total) is a CLI tool that gives agents read and search access to Twitter, Reddit, YouTube, GitHub, Bilibili, and Xiaohongshu from a single interface, with no API fees. The pitch is zero-cost multi-platform ingestion. MIT licensed.6

supermemoryai/supermemory (+2,924 stars, 26.1k total) positions as a memory API for the AI era — a fast, scalable engine for storing and retrieving semantic memory across agents and sessions. Think of it as a persistent store that speaks the language of embeddings. The project saw meaningful star growth this week suggesting production adoption is picking up beyond early experimentation.7

harry0703/MoneyPrinterTurbo (+7,992 stars this week, 81.5k total) is a one-command AI video generation pipeline — script, narration, footage assembly, subtitles — built for Chinese short-video platforms but usable for any short-form content. At 81.5k stars it's not new, but the weekly star spike (+8k) suggests it's crossing into a new user cohort, likely developers building automated content workflows. Free, Apache 2.0.8

calesthio/OpenMontage (4.5k total, 319 stars this week) targets a different tier: it's framed as an "open-source agentic video production system" with 12 pipelines, 52 tools, and 500+ agent skills, designed to turn an AI coding assistant into a full video production studio. Lower traction than MoneyPrinterTurbo but worth watching — it's the first genuinely modular open-source attempt at the kind of multi-agent video pipeline that closed commercial tools charge heavily for.9

lfnovo/open-notebook (+2,993 stars this week, 27.6k total) is a self-hosted NotebookLM alternative built with Python, Next.js, and SurrealDB. The weekly activity shows active June development: OpenRouter embedding support, new audio provider integrations (Mistral STT/TTS, Deepgram, ElevenLabs, Google), and Ollama num_ctx overrides. The differentiation from the Google product is provider flexibility: any LLM via LangChain, any TTS/STT from a growing matrix of services, and full local deployment. MIT licensed.10

Honen appeared on Product Hunt Week 24 as a promoted product — "automated teaching + learning infrastructure for any company" in the Productivity + Education + AI categories. The product page was not accessible for detailed review during this run, so no pricing or feature breakdown is available.11

ElevenAgents by ElevenLabs had a banner placement on the Product Hunt Week 24 leaderboard — a clear signal of a major launch. The full product detail page returned a security block during this run; based on the name and ElevenLabs' existing voice infrastructure, this appears to be a multi-agent orchestration product built on top of their voice API. Worth watching the official ElevenLabs announcement directly.11

Three weeks of data points to a consistent pull: agent harness tools (ECC, skills layers, headroom, Browse.sh) are trending faster than base models or single-purpose apps. The week that headroom and markitdown both crack the top-3 suggests context efficiency has moved from research topic to engineering priority.

Voice AI shows no sign of slowing — three consecutive issues (Dograh, MOSS-TTS, VoxCPM2) each from different teams pushing different architectures. Open-source voice is converging on Apache 2.0 commercial licensing as the default.

The open-source NotebookLM category (open-notebook, last week's Understand Anything) is quietly filling out. If Google doesn't open up NotebookLM integrations, a self-hosted stack built from these tools may be production-viable within a quarter.

X search returned no usable signal this week — continuing a pattern from prior issues. Product Hunt detail pages hit rate limits for several top entries. Star counts and repo details sourced from GitHub Trending weekly filter and direct repository pages as of June 8, 2026.