AI Open Source Radar — August 18, 2026

Data window: 21:46 UTC on August 16 to 21:46 UTC on August 17, 2026. The primary metric is Trendshift daily momentum (stars gained today). Cumulative stars are shown for reference only.

Today’s shape: DeepSeek Harness (dsh), published on August 13, set the fastest star-growth record in GitHub’s history — and its ecosystem took over the entire top of the daily trending board.

deepseek-ai/deepseek-harness — +13,800 ⭐ today (about 141,500 ⭐ total)

https://github.com/deepseek-ai/deepseek-harness

This is DeepSeek’s own open-source “agent harness.” A harness is the layer connecting a language model to the real world — the filesystem, shell execution, tool definitions, sessions and approvals, and long-running workflows — and dsh’s design principle is that every one of those components, without exception, is a plugin (“Everything is a Plugin”). It sits on top of the Cordis framework and exposes the following as swappable modules:

Tool definitions
Memory and context management
Sub-agent orchestration
Loop control
Stop conditions — deciding when execution halts

Installation runs on Node.js: npx @deepseek-ai/dsh web brings up a local web UI (127.0.0.1:3080 by default), and headless execution is supported. After its August 13 release it passed 20,000 stars in 90 minutes — the fastest pace GitHub has recorded — and cleared 140,000 within four days. Another 13,800 landed today, with no sign of deceleration. Worth noting: it is still a developer preview, and compatibility-breaking changes are expected.

Practical angle: If you want an in-house coding agent without vendor lock-in, you can swap tools, approval policies, and stop conditions plugin by plugin to match your organization’s rules.

Tags: #AgentHarness #PluginArchitecture #CodingAgent #SubagentOrchestration #SelfHosted #ToolCalling

anywhere-labs/deepseek-harness-desktop — +2,200 ⭐ today (about 11,100 ⭐ total)

https://github.com/anywhere-labs/deepseek-harness-desktop

A native macOS and Windows desktop client that wraps the official DeepSeek Harness in Electron. The clever part is structural: the desktop shell itself is a dsh plugin. Windows, tray, terminal, auto-update, and work configuration are all implemented as proper plugins, pushing the harness’s plugin philosophy all the way up into the UI layer. Users get a single installer that automatically starts and manages the local harness service — no Node.js install, no commands — with system tray and desktop window integrated. It also supports remote connections from iOS and Android to kick off tasks and watch agent progress, which makes checking on long-running jobs from a phone practical. Current stable builds cover Windows x64 and Apple Silicon Macs; Intel Macs are not yet supported. It is a community project rather than an official DeepSeek product, MIT-licensed, and has climbed steeply for four straight days since first hitting the trending board on August 14.

Practical angle: The fastest way to remove installation friction when rolling a local agent out to product, design, and operations staff who are not comfortable in a terminal.

Tags: #ElectronApp #DesktopClient #LocalAgentRuntime #RemoteMobileControl #TrayIntegration

awesome-dsh-plugin/awesome-dsh-plugin — +2,100 ⭐ today (about 5,500 ⭐ total)

https://github.com/awesome-dsh-plugin/awesome-dsh-plugin

A community-curated directory of DeepSeek Harness plugins. It goes beyond a link list: the rule is that every entry declares a dsh.bundle manifest and is installable directly via dsh plugin add. Entries are organized into categories so you can pick only the layer you need:

UI extensions and skins
Models and providers
Memory
Tools and capabilities
Vision and multimodal
Skills

What makes it more than documentation is the integration: the dsh web GUI has an in-harness plugin market where Settings → Plugins → Plugin Market browses the awesome-dsh-plugin.com catalog directly and installs or removes plugins into a profile. In effect, the awesome list functions as a package index. The plugin ecosystem grew fast enough in four days that several rival curation repositories appeared, and that demand shows up here as 2,100 stars in a single day.

Practical angle: If you are evaluating dsh, this is the first survey to run when deciding what to build versus what already exists — a natural starting point for defining your team’s standard plugin set.

Tags: #PluginRegistry #AwesomeList #EcosystemCatalog #PluginMarket #DshBundle

cactus-compute/needle — +436 ⭐ today (about 6,600 ⭐ total)

https://github.com/cactus-compute/needle

A 45-million-parameter foundation model built to run on tiny devices — phones, wearables, smart home hardware, and robots. It specializes in three agent tasks: tool calling, device use, and structured extraction. The entire model ships as a single 14MB binary and a full session runs in roughly 28MB of RAM. It builds on Simple Attention Network research, applies Cactus Quants’ CQ2 two-bit quantization, and bakes the result into its own inference engine. On benchmarks it trades wins with small models like FunctionGemma 270M, LFM2.5 230M, and Apple FM while being 5x to 70x smaller. Measured throughput:

Raspberry Pi 5 — about 500 tokens/sec decode
Meta Quest 3S and Apple Vision Pro — 400 to 1,500 tokens/sec
Sub-$200 phones such as the Samsung A series — 300 to 700 tokens/sec
Reported running on ESP32-P4 and ESP32-S3 class microcontrollers

Usage is pip install cactus-needle, describe your tools, and call them from Python. The inference engine is fetched once from Hugging Face and cached, and offline installation for air-gapped devices is documented.

Practical angle: For IoT, kiosk, and robotics products that need to turn a spoken command into a tool call without a cloud round trip, it removes bandwidth cost, latency, and privacy exposure in one move.

Tags: #OnDeviceInference #EdgeAI #TinyML #Quantization #FunctionCalling #Microcontroller

Excluded Today

nexu-io/open-design (+571, 87.7k total) — near the top continuously since April; growth is negligible against its base, so it counts as rank-holding rather than surging.
earendil-works/pi (+658, 92.1k total) — a trending regular since January. Mature and flat.
unslothai/unsloth (+470) and harry0703/MoneyPrinterTurbo (+573) — large, mature, permanently present.
tt-a1i/archify (+398) — already covered in the August 9 edition; momentum has slowed since.
guillaumemeyer/watermarks-remover (+1,300) — a tool for stripping C2PA and provenance marks from AI-generated content. Excluded on editorial grounds, since its clear purpose is circumventing attribution.
amirappleidfd-stack/spider–panel (+2,095), public-apis (+1,000), basecamp/omarchy (+468), liquidslr/system-design-notes (+592) — not AI tools.