AI Open Source Radar — August 26, 2026

Selection rule: recent growth is the primary signal — how much a repo gained yesterday and today. Cumulative stars are reference only. Repos that keep a high rank without accelerating are cut, no matter how large. Data comes from a three-way cross-check of Trendshift daily, Trendshift weekly, and individual repository pages.

bilawalsidhu/gods-eye-view — +1,400 ⭐ today (~1.7k total)

github.com/bilawalsidhu/gods-eye-view

An open-source spatial-intelligence platform that puts the whole planet in one browser tab. It layers live aircraft, ships, satellites, earthquakes, traffic and public camera feeds onto a photorealistic 3D globe, and the UI explicitly marks which layers are genuinely live and which are modeled, estimated or reconstructed — that distinction between real and simulated is the design principle the project leads with. A realtime AI voice agent sits on top, so you can ask hands-free questions like “what’s happening in the Strait of Hormuz right now,” have it refresh the underlying data and answer, annotate the 3D scene, or jump straight into a first-person cockpit view of a specific aircraft. It ships under MIT, so adding new data layers or extending the voice agent is fair game. Creator Bilawal Sidhu released V1 on August 24 and the pitch that spread on X — “a spy simulator in your browser, except the data is real” — drove more than 80% of the repo’s entire star count in a single day, taking the #1 spot on Trendshift’s daily list.

Practical angle: logistics, insurance and energy risk teams can self-host a situation-room dashboard and query vessel traffic, flight anomalies and disaster events by voice.

Tags: #SpatialIntelligence #ADSB #AISVesselTracking #RealtimeVoiceAgent #Geospatial3D #OSINT

FlashML-org/FreeToken — +1,400 ⭐ today (~5.9k total) · third straight day

github.com/FlashML-org/FreeToken

An inference engine for serving frontier-scale Mixture-of-Experts models on edge and consumer hardware, released by researchers at UC Berkeley and UT Austin alongside the arXiv paper FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution (2608.16157). The core idea is to stop treating GPU, CPU, host memory and interconnect as separate resources and instead manage them as one elastic inference pool. Concretely it combines four things:

bandwidth-adaptive CPU–GPU co-execution
full-layer double-buffered prefill streaming
global LRU expert caching
semantic-aware caching

The headline result is running the 753B-parameter GLM-5.2 on a single RTX PRO 6000 workstation GPU at roughly twice the throughput of llama.cpp under the same conditions. It is Apache-2.0 and published to PyPI as freetoken v0.1.2. After +757 on Aug 23 and +852 on Aug 25, it gained +1,400 today — the press wave that started with MarkTechPost and Dataconomy is accelerating rather than fading, which is rare enough to earn a third consecutive appearance.

Practical angle: for customers who cannot move data off-premises, one workstation is now enough to run a very large open-weight model in a PoC and measure real cost against a cloud inference quote.

Tags: #MoEServing #EdgeInference #CPUGPUCoExecution #ExpertCaching #BandwidthAdaptive #ConsumerGPU

AgriciDaniel/claude-obsidian — +415 ⭐ today (12.1k total)

github.com/AgriciDaniel/claude-obsidian

A plugin that turns an Obsidian vault into a self-organizing second brain maintained by Claude Code. It implements Andrej Karpathy’s LLM Wiki pattern: drop any source into the vault and Claude reads it, turns it into linked, source-cited notes, and files them into one connected knowledge graph. The structure is deliberately minimal — three kinds of plain files (a hot cache, an index, and your wiki pages), no database and no embeddings server. The README makes a point of what it refuses to do: your vault stays a normal directory of Markdown, JSON and source files, not hidden in a plugin cache, not locked in a cloud database, not silently uploaded to a model. Workflows are split into four explicit tracks — research, retrieval, maintenance, and visual mapping — and the agent answers only from evidence already in the vault, flags contradictions between sources, and keeps a living index. MIT licensed. It first hit #6 on GitHub Trending on August 24 and re-accelerated today into the #6 daily slot, running about 3.6x its prior daily average.

Practical angle: consulting and research teams can dump vendor docs and meeting notes into one folder and watch a source-cited internal wiki grow, with the output staying as vendor-neutral Markdown.

Tags: #PKM #ObsidianVault #LLMWiki #KnowledgeGraph #MarkdownFirst #CitationTracking

laude-institute/headlong — +350 ⭐ today (~0.7k total)

github.com/laude-institute/headlong

A “microharness” from the Laude Institute whose core is under 10K lines of Bash. Its defining property is persistent agency: the agent keeps thinking between external interactions in a self-guided loop modeled on human inner monologue. A message from a person does not start a session — it drifts into that thought stream as one more observation, and the agent decides whether and when to respond. Four components make it up:

a loop that generates the next thought
shellm, a recursive language model written in Bash
a trajectory stored as a DAG of jsonl files
context, defined as a projection of that trajectory

You give the agent a name and a personality; it sets its own interests and priorities, starts its own projects, and pings you when it has something to say. It is built to be shared rather than owned — a whole team can talk to one agent over Slack, Telegram and a chat app, with every conversation landing in a single stream of thought. Apache-2.0, six contributors, and today’s spike came from a Hacker News posting plus Chinese-language coverage on X.

Practical angle: for always-on operational duties — build status, inventory, ticket queues — you can try a single resident agent in a team Slack that watches on its own and speaks up first when something looks wrong.

Tags: #PersistentAgency #RecursiveLLM #BashHarness #InnerMonologue #DAGTrajectory #SharedAgent

ApodexAI/FrontierAgent — +312 ⭐ today (~0.65k total)

github.com/ApodexAI/FrontierAgent

The agent framework and execution harness Apodex open-sourced alongside its Apodex 1.1 model. It gives you a native command-line TUI and starts with a single command on macOS and Linux — no preinstall step, no hard Docker dependency. There are two execution modes: ReAct, where one agent runs research and file tasks sequentially, and Agent Team, where a coordinator decomposes the task and dispatches sub-agents in parallel. The harness handles four axes of execution — File, Search, Code and Agent Team — paired with a locally deployable 35B open-weights model, Apodex 1.1 mini. The accompanying 70-author paper proposes two scaling axes, “Environment Scaling” (expanding the set of executable code environments) and “Agentic Coordination Scaling” (parallel task delegation across a shared harness), and reports the 35B mini with Agent Team scoring 50.2 on FrontierFinance, 51.7 on FrontierScience-Research and 27.7 on APEX-Agents — effectively matching the 1T-parameter Kimi K2.6 at 27.9. The paper topped Hugging Face Papers on August 24, and more than 90% of the repo’s stars arrived today.

Practical angle: verification-heavy long-horizon work in finance or scientific research can be run on a 35B local model with parallel sub-agents, letting you weigh cost and data control against a frontier API in the same test.

Tags: #AgentTeam #ReActLoop #TerminalTUI #OpenWeights35B #ParallelSubagents #DeepResearch

Notable exclusions and why

freestylefly/awesome-gpt-image-2 — +2.2k yesterday, +989 today (-55%). Cut under the pre-committed rule that a day-two gain below +1k means exclusion.
deepseek-ai/deepseek-harness — still large at +2.2k, but short of the +2.8k bar set as its one-time re-entry condition. At 189.6k cumulative it is now permanently excluded.
Alishahryar1/free-claude-code (45.4k), DietrichGebert/ponytail (109.7k), multica-ai/andrej-karpathy-skills (205.6k), tinyhumansai/openhuman (37k+) — large, permanently trending, rank-holding rather than spiking.
wafer-ai/gpu-perf-engineering-resources and zhaoxuya520/reverse-skill — featured yesterday, both dropped out of today’s daily top 25.
b-nnett/grok-bot-0.18-reconstructed — an unofficial reconstruction of a commercial app; excluded on editorial grounds.
Non-AI subjects: openGym, vorssaint-utils, xyops, OpenLogi, walgit, Win11Debloat.