New
google-research/harness2 — research code for iterative improvement of agent harnesses
Introducing google-research/harness2, which improves tools, instructions, and memory configuration based on task results while keeping the model fixed.
CURATED OPEN SOURCE INTELLIGENCE
Latest news
New
wandermyz/vhotseat v0.1.0 — GPU Job Resource Hand-Off Tool
wandermyz/vhotseat is a Python tool that mediates so a local LLM and ComfyUI can take turns using GPU resources without interrupting work that is already running.
Release
knifelemon/filee v1.12.0 — Adds PDF Conversion and Compression
knifelemon/filee v1.12.0 adds the ability to convert PDFs into image-slide PPTX files and to compress PDFs, along with video frame extraction and hardware encoding support.
New
intel/dokimi — Measuring the Effect of Coding Agent Artifacts
intel/dokimi is a Python framework that uses paired comparison to measure whether agent artifacts such as skills and documents actually help the model.
Release
fazmin/cve-intel v0.3.4 — Local CVE Classification Tool
fazmin/cve-intel released v0.3.4, but the release body contains only a comparison link to the previous version, so the specific changes cannot be confirmed.
New
feder-cr/invisible_playwright_mcp v0.72.3 — browser automation server
feder-cr/invisible_playwright_mcp is a Python project that provides browser automation through a web interface and an MCP server.
Release
nace-ai/kapitsa v0.1.36 — Model Requirements and Benchmark Tables Consolidated
nace-ai/kapitsa v0.1.36 consolidates per-model requirements and a comparative benchmark table, and covers an M1 thread-setting issue and macOS 14 build compatibility.
Release
fazmin/cve-intel v0.3.2 — Only a Changelog Link Published
fazmin/cve-intel v0.3.2 has been released, and the release body contains only a link to a changelog comparing it with the previous version, so the specific changes cannot be confirmed.
Release
chekusu/yobiru v0.5.3 — rebuild configuration and bot instructions updated
chekusu/yobiru v0.5.3 updates the configuration and bot instructions for rebuildable Linux cloud computers, covering how to check where state is stored and how to verify messaging after a migration.
New
nace-ai/kapitsa v0.1.36 — an inference engine for Apple silicon
nace-ai/kapitsa is a C++ inference engine that runs Qwen and Nemotron-family models on Apple silicon.
Release
atoll92/agent-standup v1.39.0 — fixes shutdown warnings and call resume errors
atoll92/agent-standup v1.39.0 is an update that fixes agent session warnings and call resume errors, and describes an optional usage statistics request.
Release
opensec-git/pi-bolt bolt-v0.9.5 — request serialization and event delivery path changes
opensec-git/pi-bolt bolt-v0.9.5 reuses request serialization results in long conversations and reduces the streaming delivery path and model configuration reads.
Release
ggml-org/llama.cpp b11723 — Hexagon attention path improvements
ggml-org/llama.cpp b11723 is a release that improves FLASH_ATTN_EXT in the Hexagon backend, refining how long-context and sliding-window attention models are handled.
Release
sgl-project/sglang v0.5.22 — Model support and serving changes
sgl-project/sglang v0.5.22 is a release that brings new model support, model loading changes, disaggregated prefill/decode placement, and an expanded scope for sparse attention.
Release
tangency-labs/solidcraft v0.16.2 — Fixes Perspective-View Rotation Manipulation
SolidCraft v0.16.2 fixes navigation so that, when rotating a model in perspective view, the selected point stays under the cursor.
Release
lumarylabsllc/muxin v0.5.8 — Chinese input candidate window and worktree inspection fixes
muxin, which gathers agent sessions and desktops from multiple computers into one app, shipped ten fixes in its v0.5.8 release on GitHub, covering input and sharing features among others.
New
berriai/moyai — a self-hosted coding agent with durable execution
berriai/moyai is a self-hosted workspace that combines durable execution with isolated computers to run coding work in the background.
Release
professorpalmer/bonsai-ada-surgery bundle-20261018e — change to when the long-conversation cache is saved
bundle-20261018e leaves the inference engine binary unchanged and defers the full cache save for long conversations to idle time.
New
vectifyai/pageindex v0.4.3 — Vectorless document-tree RAG
vectifyai/pageindex v0.4.3 is a Python library that builds a hierarchical tree per document and lets an LLM follow it during retrieval.
New
deepseek-ai/deepseek-harness — a plugin-style agent harness
A look at the structure of DeepSeek's open-source plugin-style agent harness deepseek-ai/deepseek-harness and the limits of its developer preview stage.
No matching stories.