news.ycombinator.com
I made a secure way for agents to request secrets from you using HyperDHThttps://news.ycombinator.com/item?id=49346770Hi all,I kinda got sick of having to give secrets to my agents and all the potential leakage in the pipeline (with the harness, the model router, the model provider, the training set, the chat application etc etc) so I decided to make peardrop.fyi - this tool allows your agent to declaratively generate secret request pages/links which you can fill in via web or CLI. The agent can determine a script that runs once the values are received or can put them in a target folder. This is useful if you want to put something in your machine vault/keychain without either giving access to the credentials or the browser to the agent.here is the repo: https://github.com/smashah/peardrop(cli, core and self-hostable relay are all open source)Aug 18, 2026 3:08 PM
github.com
Show HN: Neal – Codex writes the code, Claude reviews ithttps://github.com/navels/nealneal is a CLI I wrote while trying to use an LLM coding agent (in this case, Codex GPT-5.4) to work autonomously on a migration of our large frontend codebase. A few things came up during that project that guided the development of neal:1. Telling an agent to "keep working unless blocked" doesn't actually work over long stretches of time.2. Large projects should be broken up into smaller chunks of work.3. The coding agent should start each chunk of work with a fresh context to prevent context rot.4. The coding agent benefits from having a different agent doing adversarial reviews along the way.What I ended up with is an orchestrator that- lets you configure planner, coder, and reviewer roles. I started with Codex and Claude using their SDKs but later added OpenRouter as well as a compatibility mode for verifying that a model can handle the orchestration. Currently there are 44 compatible OpenRouter models.- runs a plan that you provide through a planner / reviewer loop which splits theAug 4, 2026 1:14 PM
github.com
Show HN: Open-Cowork – an open-source, model-agnostic computer-use agenthttps://github.com/coasty-ai/open-coworkHi HN,We built Open Cowork, an MIT-licensed desktop agent that can operate a computer through screenshots, mouse input, and keyboard input.We started working on this after using cowork-style computer agents for longer tasks. The interaction model was useful, but computer use burns through model calls much faster than chat: take a screenshot, decide what to do, perform an action, inspect the result, and repeat.We wanted to separate that agent loop from any one model or subscription.Open Cowork currently works with local models through Ollama and hosted models through providers including OpenAI, Anthropic, Google, xAI, Mistral, and OpenRouter. You can watch the agent’s actions as they happen and interrupt it during a run.The basic loop is:Capture the current screen Send the screenshot and task state to the model Parse the proposed action Execute the mouse or keyboard action Capture the resulting state Continue until completion, interruption, or a safety conditionTwo annoyingly common faiJul 30, 2026 5:54 PM
news.ycombinator.com
I Replaced My AI Agent's Flat Fact Store with a Graph Databasehttps://news.ycombinator.com/item?id=48383578# I Replaced My AI Agent's Flat Fact Store with a Graph Database and It Runs in 85MBI've been building LocalClaw, a local-model-first AI agent framework running on personal hardware through Ollama. No cloud, no API costs. A few weeks ago I posted about the router/specialist architecture. A lot of people asked about the memory system so here's that.## The ProblemStarted with a JSONL fact store and embedding similarity retrieval. Simple enough until it wasn't. After a few weeks of real use I had 14 near-duplicate facts about the same topics from different sessions. Layered dedup on top of dedup and it still wasn't clean.The bigger problem was relationships. "Peter works at DevMesh" and "DevMesh is building an outreach platform" were two separate embeddings. You could retrieve each one but you couldn't traverse from one to the other. No multi-hop. No fact evolution. Old facts and new facts coexisted with no signal about which was current.Four iterations on the flat store later I accepted Jun 3, 2026 1:15 PM
github.com
Show HN: Monkdev is a toolkit and methodology for coding with LLMshttps://github.com/oeo/monkdevI'm sure many people have some solutions similar to this but I've been using some variation of this since roughly the release of Opus 3 to get higher quality results. Like, significantly higher. So I decided to make it into a more structured package.Problem: bad decisions and short sighted solutions due to the LLM not having enough context before acting. They'd read one file, or a few segments of a set of files using a super conservative offset and limit, guess about the rest, apply a patch, run a unit test and then tell you with its entire chest that the code is production-ready.Those bandaids over time became debt, drift, complexity. Unnecessary LOC in a codebase where I find myself wondering is this thing even making me more efficient or should I be in VIM more often and just save myself the API costs (I use OpenCode + OpenRouter and have been known to blast through hundreds of dollars a day in spend using Frontier models).I found that, for whatever reason, LLMs respond to role-playMay 26, 2026 7:08 PM
news.ycombinator.com
Show HN: Chat – Another open-source chat UI for MCP Serverhttps://news.ycombinator.com/item?id=47398628i built this because i’ve been experimenting with MCP and felt there was a need lightweight version for ecosystem.most existing clients felt a bit too noisy and bundled into larger ecosystems, so i wanted simplest/ near plain version that just focuses on end-user chat experience.stack(main): NextJS(App Router) + Vercel AI SDKcapabilities:+ set single MCP endpoint and brand mode(bring your own brand asset) on ENV+ prepare local db or external endpoint(SQLite, PostgreSQL, MySQL)+ choose in-memory(default) or external/ local Redis endpoint+ built-in auth via Better Auth & Resend(optional)+ support file upload with Cloudflare R2+ support location-sharing with LeafletJS and Browser Geolocation or Google Maps Platform+ automatic data cleanup with Trigger(free-tier trigger.dev is ENOUGH)+ auto-detect user location so they dont need to set timezone manually+ locale currently support 10 languages(EN, ID, KR, JP, ES, ZH, DE, NL, FR, IT)+ plain/minimalist interface+ install with single interactivMar 16, 2026 1:17 PM
mindweave.space
Show HN: Mindweave – AI-powered personal knowledge hub with semantic searchhttps://www.mindweave.space/Hi HN,I built Mindweave to solve a problem I kept running into — I'd save bookmarks, notes, and links across different apps, then never find them again when I actually needed them.Mindweave lets you capture notes, links, and files in one place. The interesting part is what happens after: - Semantic search — find content by meaning, not just keywords. "That article about improving deep work" finds it even if those words don't appear in the content. Powered by pgvector cosine similarity on Gemini embeddings. - AI auto-tagging — Gemini generates tags on save, so you don't have to organize anything manually. - Knowledge Q&A — ask questions about your saved content using RAG. Retrieves relevant pieces, feeds them as context to Gemini, returns a grounded answer. Stack: Next.js 15 (App Router, Server Actions), PostgreSQL 16 + pgvector, Google Gemini (text-embedding-004, 768d), Drizzle ORM, Auth.js v5, Tailwind/shadcn. Deployed on Cloud Run. A few things I found interesting while building thisFeb 16, 2026 5:20 AM
karatsidhu.gumroad.com
Show HN: Warden – A Native (and Free) AI Chat App for macOShttps://karatsidhu.gumroad.com/l/wardenWarden is a minimalist, simple and beautiful macOS AI chat app, that supports most AI providers: ChatGPT, Anthropic (Claude), xAI (Grok), Google Gemini, Perplexity, Groq, Local LLMs through Ollama, OpenRouter, and almost any OpenAI-compatible APIs.Completely Native app, built with SwiftUIZero telemetry, no data collected from you, all your data remains on your Mac and can be easily exported into an easy-to-read JSON fileLightweight app, compared to Electron based alternativesSupports these LLM APIs:OpenrouterOpenAIAnthropic (Claude)xAI (Grok)GroqPerplexityOllamaDeepseekIt should also support most other OpenAI compatible APIshttps://karatsidhu.gumroad.com/l/wardenMay 29, 2025 6:19 AM
