AI Open Source Radar — August 3, 2026

Data window: UTC 2026-08-01 21:08 to 2026-08-02 21:08. Primary signal is recent growth (stars today), using Trendshift’s daily ranking (freshness verified) cross-checked with OrangeBot’s daily list. findarepo was excluded this cycle because its snapshot was stuck on the previous day. Large, mature, or plateauing repos — and yesterday’s picks whose growth has cooled — were left out.

yc-software/qm — +286 ⭐ today (≈3.1k total)

qm is a “multiplayer agent harness” for teams, and it jumped to the day’s #1 spot after Y Combinator open-sourced (MIT) the very tool it runs internally across its accounting, legal, events and engineering teams. Each person and each room (channels, group messages, projects) gets its own scoped memory, file system, keychain view, permissions, cron jobs, web apps and a durable sandbox, so many people can collaborate with the agent without stepping on each other. The agent acts as the person it works for — using their credentials and permissions — and every action is audited, while the harness and model are swappable across Pi, OpenCode, Codex and Claude Code, so a deployment isn’t locked to any single vendor. Security posture is chosen per org, and it’s the combination of team-wide, always-on collaboration plus the “the tool YC uses to run itself” story that drove the surge.

Strict — every harness tool call pauses for human approval (except the two no-effect turn-enders).
Auto (default) — a classifier screens provenance-labelled external data and tool results before they reach the model.
Dangerous — no content screening and no pauses between tool calls.

Practical use: With per-department isolated sandboxes and audit logs, it’s a good fit for deploying agents team-by-team in security- or compliance-sensitive organizations.

Tags: #AgentHarness #MultiAgent #SlackBot #SandboxIsolation #ToolCallAudit #VendorAgnostic
Reference: github.com/yc-software/qm

firecrawl/pdf-inspector — +62 ⭐ today (≈1.4k total)

Built by Firecrawl, pdf-inspector is a pure-Rust library that inspects, classifies and extracts text from PDFs — and, crucially, intelligently distinguishes scanned vs. text-based PDFs to make smart downstream routing decisions. By sampling content streams it classifies a document as TextBased, Scanned, ImageBased or Mixed in roughly 10–50ms, returning a confidence score (0.0–1.0) and per-page OCR routing, and it does position-aware extraction with font and X/Y coordinates plus automatic multi-column reading order. Its Markdown conversion handles headings (H1–H4 via font-size ratios), bullet/numbered/lettered lists, monospace-detected code blocks, rectangle- and heuristic-based tables, bold/italic, URL linking and page breaks. It runs with no ML models or external services — a single dependency on lopdf — and ships Python, Node.js and browser WebAssembly bindings, which is why it’s drawing attention as a cheap pre-processor that decides “when to run OCR” in RAG and document pipelines.

Practical use: Placed at the front of a RAG pipeline ingesting large PDF volumes, it can route only scanned files to OCR, sharply cutting processing cost and latency.

Tags: #PDFParsing #DocumentClassification #OCRRouting #RustLang #RAGPipeline #WASM
Reference: github.com/firecrawl/pdf-inspector

QwenAudio/qwen-audio-agent — +35 ⭐ today (≈561 total)

qwen-audio-agent is a realtime voice runtime designed to keep agents “talking, working, and present.” Its core is full-duplex realtime voice interaction with natural interruption and continuous multi-turn conversation, and you can pick your preferred agent with one click and reuse its existing tools, MCP servers and Skills. Conversation keeps flowing while the agent works in the background, and when the Agent Protocol is empty the gateway falls back to a frontend-only mode where realtime voice chat still works. The desktop app provides a persistent “voice orb” and embeds and manages its own gateway, so no service needs to be started in advance; on first launch it creates a config file and guides you to enter a DashScope API key and choose a backend agent (Node.js 22.22.2+/24.15.0+ required). It’s riding demand to bolt Alibaba’s Qwen audio stack onto agents as a voice frontend.

Practical use: Add a hands-free voice interface to an existing coding or work agent for on-the-go, multitasking consultation and command scenarios.

Tags: #VoiceAgent #FullDuplex #RealtimeVoice #MCP #DashScope #VoiceInterruption
Reference: github.com/QwenAudio/qwen-audio-agent

TencentCloud/TencentDB-Agent-Memory — +31 ⭐ today (≈386 total)

This is a team-level memory hub that turns conversations, docs and code into four reusable memory assets that are governed, shared and equipped across agents and frameworks. Import a codebase and Code-Graph automatically indexes symbols, files, call relationships and impact paths; docs and files become structured Wiki pages with a link graph; and past agent sessions have Skills and Chat Memory automatically extracted. Its design philosophy rejects both brute-force history accumulation and irreversible lossy summarization — handling in-task information overload with symbolic memory and cross-session experience with layered memory. Notably, it preserves a deterministic drill-down path from top-layer abstractions back to bottom-layer raw evidence for traceability, and delivers fully local long-term memory via a 4-tier progressive pipeline with zero external API dependencies.

Chat Memory — extracted from conversation sessions.
Skill — reusable patterns from past interactions.
LLM-Wiki — structured documentation with a link graph.
Code-Graph — indexed codebase symbols and relationships.

Practical use: Let a team’s multiple agents share the same code, docs and know-how, reducing the cost of rebuilding context during onboarding and repetitive tasks.

Tags: #AgentMemory #CodeGraph #LongTermMemory #LocalFirst #KnowledgeGraph #SkillExtraction
Reference: github.com/TencentCloud/TencentDB-Agent-Memory

sqliteai/waste — +23 ⭐ today (≈367 total)

WASTE (Weight-Aware Streaming Tensor Engine) is a dependency-free, embeddable C inference engine that runs the full 2.78-trillion-parameter Kimi K3 model beyond available RAM by streaming activated expert weights directly from NVMe. Kimi K3 is a Mixture-of-Experts model where only a small fraction of its 896 experts per layer activates per token, so WASTE keeps the dense, repeatedly used trunk resident in memory, stores routed experts in an NVMe-optimized container, streams only the experts selected during inference, and uses remaining RAM as a bounded expert cache. It currently runs the 1.42TB of weights (a 982GiB container) on a 64GB MacBook Pro at about 0.6 tokens per second. As a Show HN that proves frontier-scale models can actually run — if slowly — on consumer hardware, it’s riding the same local mega-model inference wave as projects like turbo-fieldfare and jcode.

Practical use: Useful for research and privacy-sensitive settings that need to validate or prototype giant open-weight models locally without expensive GPUs.

Tags: #LocalLLM #MoE #NVMeStreaming #KimiK3 #TrillionParameter #CInference
Reference: github.com/sqliteai/waste

Selection criteria: new, smaller repos with the steepest growth within the last 24 hours (UTC). Large, well-known perennial repos (e.g., OmniRoute, orca, superpowers, graphify) and yesterday’s picks whose growth cooled (turbo-fieldfare, quill, aos-ce, SimpleEnglish) were excluded. Dual-use security (reverse-engineering / penetration) repos were deliberately left out even when surging.