Category / T-CAT / SOFTWARE
Software
Software / T-2026-9928
Karpathy's coding critique, packaged as a CLAUDE.md file
A CLAUDE.md file distills Karpathy's LLM coding critique into four principles for Claude Code.
Sourcemultica-ai/andrej-karpathy-skills
Software / T-2026-7416
Apache Maka makes the agent log the runtime, not the model
Apache Maka's local-first agent workspace treats the event log as the runtime, a design worth watching for AI accountability.
Sourceapache/maka
Software / T-2026-7434
llama.cpp b10538: the KV cache gets a batch-size switch
llama.cpp b10538 turns off KV-cache dequantization for small batches on Apple Silicon, a small change with a big signal.
Sourceggerganov/llama.cpp b10538
Software / T-2026-2296
Tencent's AI-Infra-Guard turns agent skills into a CVE-style attack surface
Tencent's open-source AI-Infra-Guard brings CVE-style scanning to agent skills, MCP servers, and LLM infrastructure.
SourceTencent/AI-Infra-Guard
Software / T-2026-3620
Anthropic's SDK update signals the real agent battleground: memory and tooling
Anthropic's v0.125.0 SDK adds managed-agent web search and sandbox memory. A small changelog, a big signal about where agent infrastructure is heading.
Sourceanthropics/anthropic-sdk-python v0.125.0
Software / T-2026-7339
llama.cpp b10506: the quiet art of dequantizing q8_0 faster on Metal
llama.cpp b10506 ships a Metal q8_0 dequantization win. The real signal is how much frontier inference still lives in kernel details.
Sourceggerganov/llama.cpp b10506
Software / T-2026-8490
OpenViking turns agent memory into a filesystem, and cuts token spend by up to 91%
Volcengine open-sources OpenViking, a context database that unifies agent memory, RAG, and skills under a viking:// filesystem.
Sourcevolcengine/OpenViking
Software / T-2026-8976
ai-memory: the agent handoff layer that treats context as a git repo
ai-memory gives coding agents a shared, persistent wiki for cross-vendor handoffs. It's a bet that context, not compute, is the real moat.
Sourceakitaonrails/ai-memory
Software / T-2026-6992
CLI-Anything's bet: every GUI app gets a terminal, and agents get the keys
HKUDS/CLI-Anything generates agent-native CLIs for desktop apps. The registry's growth signals a shift in how AI agents will control software.
SourceHKUDS/CLI-Anything
Software / T-2026-5816
Unsloth's desktop app turns local AI into a consumer product
Unsloth ships a desktop app for running and training LLMs locally, with agent hooks and remote access.
Sourceunslothai/unsloth
Software / T-2026-2780
holaOS bets agent portability beats model loyalty
holaOS is an open-source agent workspace that runs Claude Code, Codex, or its own agent over one shared memory and 100+ integrations.
Sourceholaboss-ai/holaOS
Software / T-2026-9802
Cactus Compute's 14MB Needle 2 shrinks tool-calling AI to fit on a phone
Cactus Compute ships Needle 2, a 45M-parameter tool-calling model in a 14MB binary that runs in 28MB of RAM.
Sourcecactus-compute/needle