Research & Knowledge

The author opened a scoped connector-migration issue in the okou-ai/okou repo, as a native child of a larger HTTP-to-MCP shortlist (#36738). They identified the existing `algolia`

The author built the first half of a learning loop for a multi-harness agent server: the switch, the folder, and the seed every run starts from. Concretely: (1) a `learning: boolea

They wanted a production-grade inference path for a mixture-of-experts model on Hopper and Blackwell GPUs, and started from FlashInfer because it exposes unified APIs for attention

The author built SecondBrain 2.0 as a large draft PR on top of a Markdown/Git knowledge base, keeping Markdown and Git authoritative while layering services on top. The wiki is the

They built little-coder, a coding agent scaffold layered on top of pi (a minimal agent substrate with an agent loop, multi-provider API, TUI, session tree, compaction, and extensio

They had a LangGraph research agent that could search the web, read internal docs, and write summaries to a shared knowledge base. Before shipping it, they wanted guardrails: the a

The author noticed that the `trace_crash` prompt still recommended complete function dossiers even after `build_call_path` had gained retained-parent references (#8435, fixed in #8

They installed Claude Octopus alongside Claude Code and kept the ordinary Claude-native path for routine work, escalating only when they wanted a second opinion. For a risky refact

They set up CUAWright (formerly Webwright) as a terminal-driven harness so a coding model could launch Playwright browser sessions, inspect pages, and write scripts instead of pred

The team needed to serve large MoE models (DeepSeek-R1 671B, Qwen3, GLM-4.5) in production but did not want to hand-assemble a serving stack from scratch. They picked Chitu (赤兔), a

They took the Vibe-Trading repo and stood up a personal trading agent rather than a one-off script. Starting from the pip-installable `vibe-trading-ai` package (Python 3.11+, FastA

They set up vLLM inference on Huawei Ascend hardware using the vllm-ascend plugin, which extends upstream vLLM with Ascend NPU device support. The work followed the official docume

They wanted a self-hosted, OpenAI-compatible inference endpoint for an internal chat and summarization feature, without sending data to a hosted API. Starting from the LMDeploy REA

They started from the Langfuse repo to get a self-hosted observability backend, then wired their existing LlamaIndex retrieval pipeline into it so every query produced a trace: ret

They wanted a production-grade way to serve PaddlePaddle LLM and VLM checkpoints without rewriting their existing inference clients. Starting from the FastDeploy README, they picke

They installed Atomic Agent, a local-first agent that keeps its control loop and state on the machine, and pointed it at a llama.cpp backend so small quantized models could handle

They used Hermes One (hermes-desktop) as the front end for a self-hosted Hermes Agent instead of hand-managing the CLI. The app ran the official Hermes install script, placed Herme

They installed Attention Span, a small collection of Claude Code output styles, to change how the agent talks rather than how it codes. Each style is a single markdown file dropped

A computational biology group wanted to standardize how their lab turns raw 10x Genomics count matrices into annotated cell-type calls. Instead of hand-writing another bespoke note

They started from a project root and ran `npx impeccable install`, reloaded their coding tool, and ran `/impeccable init` so the agent read the codebase and wrote a PRODUCT.md capt

They installed the Eigent desktop app, which wraps CAMEL multi-agent orchestration in a local cowork interface. Instead of writing agent code, they described a task in natural lang

They wanted their coding agent to stop grepping blindly and re-reading whole files, so they stood up XERJ, a single local Rust binary that indexes a repository's code, docs, specs,

They wanted agents to stop guessing about a mid-sized Python/TypeScript monorepo, so they ran Graphify over the repository to extract a structural graph of modules, imports, and ca

They installed AutoHarness as a Claude Code plugin from the tigerless-labs marketplace, then let it run in the background while doing normal work. Each session that did enough work

They installed the headroom-ai package and used `headroom wrap claude` and `headroom wrap codex` to route existing agent sessions through the local compression proxy on port 8787,

They started from a plain markdown changelog and a rough product brief sitting in the repo. Instead of hand-editing a doc, they pointed html-anything at the repo and let it detect

The author (bojieli) published a full book, 深入理解 AI Agent：设计原理与工程实践, built around the formula Agent = LLM + context + tools. The repo ships the complete manuscript, figures, and 10

They took the llama.cpp WebUI and turned it into a friendlier desktop-style chat app. The fork keeps the llama.cpp server as the model backend but layers on a redesigned interface

They cloned the AgenticSeek repository and stood up a fully local, voice-enabled agent stack on their own machine. Instead of pointing at a cloud API, they wired the assistant to a

They installed the planning-with-files skill globally via the Agent Skills CLI and let its lifecycle hooks drive a three-file planning loop inside a repo. Instead of relying on a p

The developer wanted to stop paying per-token API bills while prototyping an internal chat assistant, so they self-hosted an open-source model instead of calling a hosted provider.

They started from the microsoft/generative-ai-for-beginners curriculum, working through the early lessons on generative AI and LLMs, prompt engineering, and building applications w

They stood up invisible_dots on a machine with hardware virtualization, ran `setup --all` to install QEMU and build the Dot images, and opened the local web UI at 127.0.0.2:3000 wi

They wanted to know whether their agent stack's inference layer was actually the bottleneck, so they built a benchmark harness around TensorRT-LLM instead of trusting single-reques

They took the obsidian-second-brain repo, which ships 8 agents and 48 commands, and pointed several different coding CLIs at the same Obsidian vault so the vault, not the CLI, hold

They used career-ops, the open-source AI job search agent, as their primary filter during a real search. Instead of blasting CVs into silence, they pasted job listings into the age

They started from a rough, minimal role prompt (for example a reviewer or a marketplace bargaining reply) and ran it through Prompt Optimizer's web app to push a small model away f

They wanted local inference that a teammate could run without a Python environment, CUDA toolkit, or model-download dance. Starting from the llamafile README, they downloaded a sma

The maintainers assembled an opinionated, stack-agnostic catalog of 107 Claude Skills that covers the full lifecycle of building, launching, and growing a brand and website: brand

They stood up Ekko Studio as a self-hosted web console (npm global install, then `ekko-studio-web start`) rather than the desktop build, so the backend would live on the same machi

The user wanted to add a small but real feature to an existing TypeScript web app: a saved-search feature with a REST endpoint, a UI toggle, and persistence. Instead of prompting t

They wanted a repeatable way to turn a rough question into a polished written brief without babysitting a chat window. Starting from the AutoGen quickstart, they installed autogen-