标签：#mcp

클로드로 한글파일(HWP) 변환·자동화하는 법 2026 — 요약·표 추출·일괄 처리 실전 한글파일을 Claude로 다루려는 한국 기업 실무자가 가장 먼저 부딪히는 벽은 " 읽기는 됐는데, 그래서 뭘 어떻게 자동화하지? "다. HWP-MCP를 설치해 Claude가 한글 문서를 읽게 만드는 것까지는 HWP-MCP 도입 가이드 에서 다뤘다. 이 글은 그 다음 단계 — 실제 업무에서 한글파일을 요약·변환·일괄 처리하는 구체적 방법 을 실전 예시로 보여준다. 한글파일 AI 자동화의 핵심은 "한컴 오피스 라이선스 없이, 사람 손을 거치지 않고, 반복 작업을 Claude에게 위임하는 것"이다. 계약서 100건 요약, 요구사항서의 표를 CSV로 추출, 폴더 안 HWP 일괄 변환 — 이런 작업이 자동화 대상이다. 한글파일 자동화로 풀 수 있는 업무 3가지 업무 수동 작업 시간 자동화 후 적용 키워드 문서 요약 1건당 10~15분 50건 30초 claude 한글파일 요약 표 → 데이터 추출 1표당 5분 (재입력) 표 자동 CSV 변환 hwp 표 추출 일괄 변환·정리 100건 8시간 100건 1시간 20분 한글파일 일괄 처리 세 업무 모두 "사람이 한글파일을 열어 읽고, 내용을 옮겨 적는" 반복 작업이다. Claude + HWP-MCP 조합은 이 중간 단계를 없앤다. 전제: HWP-MCP 연결 확인 자동화에 들어가기 전, Claude가 한글파일을 읽을 수 있는 상태인지 확인한다. (설치 절차는 HWP-MCP 도입 가이드 참조.) # Claude Desktop 설정에서 hwp-mcp 서버가 연결됐는지 확인 # MCP 도구 목록에 hwp_read, hwp_extract_tables 등이 보여야 함 연결이 확인되면 아래 3가지 워크플로우를 바로 쓸 수 있다. 워크플로우 1: 한글파일 요약 자동화 계약서·보고서·요구사항서처럼 길이가 긴 한글 문서를 Claude에게 요약시키는 패턴이다. 단일 문서: "이 한글파일을 읽고 다음 3가지로 요약해줘: 1. 핵심 내용 5줄 2. 의사결정이 필요한 항목 3. 누락되거나 모호한 조항" 여러 문서 일괄 요약: 폴더 경로를 주고 "이 폴더의 모든 .hwp 파일을 각각 위 형식으로 요약하고, 결과를 하나의 마크다운 표로 정리해줘"라고 지시하면, Claude가 HWP-MCP로 파일을 순회하며 처리한다. 50개 문서 기준 약 30초. 요약 품질을 높이는 팁: "요약 기준"을 구체적으로 명시 할수록 결과가 좋다. "계약 금액·기간·위약 조항 중심으로" 같은 도메인 컨텍스트를 주면 일반 요약보다 실무 적합도가 크게 오른다. 워크플로우 2: 표 → CSV 데이터 추출 한글파일의 표는 복사-붙여넣기로 옮기면 서식이 깨지는 게 가장 큰 골칫거리다. HWP-MCP의 표 추출 기능을 쓰면 구조를 유지한 채 데이터만 뽑는다. "이 한글파일에 있는 모든 표를 추출해서 CSV로 변환해줘. 표가 여러 개면 각각 별도 파일로, 헤더 행을 포함해서." 활용 시나리오: 견적서·정산표 : 한글 견적서의 항목·단가·합계를 회계 시스템에 올릴 CSV로 요구사항 명세 : 기능 목록 표를 이슈 트래커(Jira/Linear) import 형식으로 설문·조사 결과 : 한글 보고서의 통계 표를 분석용 데이터프레임으로 표 안에 병합 셀이 있으면 Claude에게 "병합 셀은 상위 값으로 채워줘(forward fill)"라고 미리 지시하는 게 데이터 정합성에 좋다. 워크플로우 3: 폴더 일괄 처리 가장 ROI가 큰 패턴. 수백 개 한글파일이 쌓인 폴더를 통째로 처리한다. "./contracts 폴더의 모든 .hwp 파일에 대해: 1. 계약 상대방·금액·시작일·종료일을 추출 2. 하나의 CSV로 통합 (파일명을 첫 열에) 3. 종료일이 30일 이내인 계약은 ⚠️ 표시" 100건 기준 수동 8시간 작업이 약 1시간 20분으로 줄어든다(실측). 핵심은 추출 스키마를 먼저 정의 하는 것 — 무엇을 뽑을지 명확할수록 일괄 처리 정확도가 높다. python-docx·한컴 API와 무엇이 다른가 방식 한글파일(.hwp) 지원 자동화 난이도 AI 통합 한컴 오피스 자동화

Claude Code에서 Codex로 옮길 때 — 실전 마이그레이션 가이드 2026 Claude Code 기반 워크플로우를 OpenAI Codex CLI로 옮기려는 팀이 늘고 있다. 모델 가격, 멀티 벤더 리스크 분산, 특정 코딩 워크로드의 성능 차이 등 이유는 다양하다. 그런데 두 도구는 같은 "AI 코딩 에이전트"라는 카테고리에 속해도 컨벤션·확장 메커니즘이 다르다. 무작정 옮기면 자동화 파이프라인의 절반이 깨진다. 이 가이드는 Claude Code → Codex 마이그레이션을 실제로 끝내본 팀이 어떤 순서로 무엇을 옮기고, 무엇을 포기하고, 무엇을 대체했는지 정리한다. 자동 변환 툴( claude2codex )을 어디서 쓰고 어디서 안 쓰는지, 일주일 점검 체크리스트, 양쪽을 분기 사용하는 하이브리드 패턴까지 다룬다. 마이그레이션 전 의사결정 — 옮길지 말지부터 옮기는 게 모두에게 정답은 아니다. 다음 세 질문에 모두 "예"여야 본격 마이그레이션을 권한다. 현재 Claude Code 비용의 60% 이상이 일상적인 코드 편집·리뷰에서 발생하는가? (Codex의 GPT-5-codex가 단가 우위를 보이는 영역) — 만약 디자인·기획·문서 분량이 큰 워크플로우라면 Claude를 유지하는 게 합리적이다. Skills·Hooks·서브에이전트 같은 Claude 고유 기능에 의존하지 않는가? 의존도가 높다면 마이그레이션 비용이 비용 절감을 초과한다. 하나의 벤더 락인을 줄이는 게 중요한 전략적 우선순위인가? 멀티 벤더 운영은 그 자체로 관리 비용이 든다. 세 질문 중 하나라도 "아니오"라면, 통째 마이그레이션 대신 하이브리드 분기 사용 (아래 5절)이 더 낫다. Claude Code와 Codex의 핵심 차이 비교 영역 Claude Code OpenAI Codex CLI 마이그레이션 난이도 메인 모델 claude-opus-4-7 / sonnet-4-6 / haiku-4-5 GPT-5 / GPT-5-codex / o1 계열 낮음 (모델 교체) 컨벤션 파일 CLAUDE.md AGENTS.md (멀티 벤더 표준) 낮음 (rename + 어조 조정) 확장 메커니즘 Skills (markdown SKILL.md + 메타데이터) 별도 표준 없음, 수동 컨텍스트 로딩 높음 (가장 큰 갭) 자동화 훅 Hooks (PreToolUse, SessionStart, UserPromptSubmit 등) 라이프사이클 이벤트 미지원 높음 (외부 wrapper 필요) 슬래시 커맨드 /명령 형태 + 인자 파싱 CLI 인자로 대체 중간 MCP 서버 1급 지원, 자동 도구 노출 일부 지원, 설정 형식 다름 중간 서브에이전트 Agent tool (subagent_type) 외부 오케스트레이션 필요 높음 권한 모드 acceptEdits / plan / dontAsk 등 --auto / --confirm 류 낮음 가장 큰 갭 세 곳: Skills · Hooks · 서브에이전트 . 이 세 가지에 깊이 의존하는 팀은 마이그레이션 ROI가 마이너스로 나올 수 있다. 마이그레이션 절차 — 7단계 1단계: 자산 인벤토리 (1일) .claude/ 디렉토리, CLAUDE.md , 프로젝트 루트의 slash command 정의, hook 설정, MCP 서버 목록을 전부 추출한다. find . -path "*/.claude/*" -type f > migration/inventory.txt ls .claude/skills/ .claude/hooks/ .claude/commands/ 2>/dev/null >> migration/inventory.txt cat .claude/settings.json | jq '.mcpServers // {}' > migration/mcp.json 이 파일들이 모두 변환되거나, 대체되거나, 폐기되는지 명시적으로 매핑되어야 한다. "그냥 옮기면 되겠지"는 거의 항상 일주일 후 장애로 돌아온다. 2단계: claude2codex 자동 변환 적용 (반나절) 오픈소스 claude2codex 마이그레이션 툴 이 자동으로 처리하는 것: CLAUDE.md → AGE

Most scraper demos lie by accident. They show the happy path: one URL, one clean page, one neat JSON object. Then the first real user tries a marketplace search page, a login wall, a JavaScript shell, a rate-limited product page, or a site that serves different HTML to every fetch path. The response still comes back as JSON, so everyone relaxes. That is the trap. A JSON response is not the same thing as a useful extraction. The failure mode agents hate AI agents do not just need scraped text. They need to know what happened. Bad extraction output looks like this: { "title" : "Example product" , "price" : "$29.99" , "availability" : "in stock" } That looks fine until you inspect the source and discover the page was a login prompt, a bot challenge, or a thin JavaScript shell. The extractor filled the schema because the schema was requested. Helpful. Like a smoke alarm that hums a little song while the kitchen burns. Better extraction output separates the data from the confidence and the failure class: { "status" : "failed" , "failure_type" : "login_required" , "confidence" : 0.94 , "extracted" : null , "evidence" : { "final_url_type" : "restricted_page" , "visible_content" : "login prompt" , "structured_data_found" : false }, "next_step" : "Use an authorised source, public item URL, feed, API, or sample HTML." } That is less flashy. It is also much more useful. The useful contract is not “scrape anything” “Scrape anything” is usually a warning label wearing lipstick. For agent workflows, the better contract is: Return structured data when the page provides enough evidence. Return a specific honest failure when it does not. Preserve enough metadata for the caller to decide what to do next. Never invent fields just because a prompt asked nicely. This matters for ecommerce, lead enrichment, price monitoring, competitor tracking, procurement, and internal research agents. If the agent cannot tell the difference between “product unavailable”, “page blocked”, “login require

If you run a website that serves European users — or if you’re based in the EU yourself — hosting location is no longer just a technical detail. It has become a strategic and legal consideration. In 2026, questions around data residency, GDPR compliance, and digital sovereignty are front and center for businesses of all sizes. Choosing where your website and its data live can affect everything from legal risk to user trust and even search engine performance. The core issue: Where does your data actually live? When you choose a hosting provider or CMS, your content, user data, images, and analytics often get stored on servers in a specific country or region. This matters because different jurisdictions have different laws regarding data access, privacy, and government requests. Key concerns include: GDPR compliance : The EU’s strict data protection regulation requires that personal data of EU citizens be handled with strong safeguards. Transferring data outside the EU/EEA can trigger additional legal requirements (such as Standard Contractual Clauses). Schrems II and data transfers : Following court rulings, simply relying on old mechanisms for transferring data to the US became more complicated. Many organizations now prefer to keep data within the EU to reduce legal complexity. Data sovereignty : Some companies and public sector organizations want (or are required) to ensure their data is stored and processed under EU jurisdiction. Performance and latency : Hosting closer to your users generally means faster load times — which also helps SEO and user experience. For many European businesses, especially those handling any form of personal data (even something as simple as contact forms or newsletter signups), EU-based hosting has become the safer, simpler default. Why traditional hosting often falls short Many popular CMS platforms and hosting providers are headquartered outside Europe. While they may offer data centers in the EU, the company’s legal jurisdiction, d

Most multi-agent frameworks for software development organize agents around roles : a product manager agent, a developer agent, a tester agent. ChatDev and MetaGPT pioneered this approach, and it works well for monolithic tasks. But I ran into a wall when I tried to apply it to a real system with multiple independently-deployed services. The Problem with Role-Based Coordination Imagine you have a backend search service and a frontend management console. The backend team implements a new API endpoint. The frontend needs to adapt. In a role-based framework, there's no natural mechanism for this. Both agents are "developers" in the same simulated organization. There's no concept of service boundaries, no versioned contracts, no way to say "the backend changed, and the frontend needs to know exactly what changed." The coordination problem in multi-service development isn't "which role should handle this task" — it's "which service needs to know about this change, and what exactly changed." That reframing led me to build something different. AgentNexus: Coordinating Agents at the Service Granularity AgentNexus is a document exchange center that treats each service as a first-class citizen. Instead of roles, it uses service boundaries as the coordination primitive. Here's how it works: Each service registers as a sub-project with its own document namespace Services publish versioned Markdown documents: requirements, design specs, API docs, config Services subscribe to documents from other services they depend on When a subscribed document changes, the subscriber receives a diff-aware notification containing both the structured diff and the full latest content The whole thing is exposed as an MCP (Model Context Protocol) server running in streamable-HTTP mode, so multiple agents can connect simultaneously from different machines. The Diff-Aware Update Protocol This is the part I'm most proud of. When an agent calls get_my_updates_with_context , it gets back: { "update_id"

I do a fair amount of competitor backlink research, and the workflow always annoyed me: open a dashboard, run a query, export a CSV, eyeball it, copy domains into a doc, switch to email. Lots of tab-hopping for what is fundamentally a data-filtering problem an agent should handle. So I wrapped the backlink API I'd been using into an MCP server. Now I stay in Claude Code (or Cursor, Cline, Zed, Windsurf) and just describe the goal. This is the build: the architecture, the four tools, and the one design decision I'm still not sure about. The data source The server runs on the Common Crawl hyperlink webgraph — about 4.4 billion edges across 120 million domains, published quarterly as Parquet. That matters for an MCP tool specifically: the data is open, so there's no scraped-proprietary-index liability in handing it to an agent, and the same query is reproducible by anyone. The HTTP API in front of it ( CrawlGraph ) does the heavy DuckDB work; the MCP server is a thin TypeScript stdio client over it. Keeping the server thin was deliberate — all the query cost, caching, and quota logic lives server-side, so the MCP package stays a ~300-line wrapper that's easy to audit before you hand it your API key. The four tools backlinks → referring domains for a target, with authority scores gap_analysis → domains linking to your competitors but not to you gap_outreach_targets → the composite play (below) releases → list the Common Crawl snapshots backlinks and gap_analysis map 1:1 to API endpoints. gap_analysis is the interesting primitive: submit your domain plus 2-5 competitors, and it returns every domain that links to at least one competitor but not to you, each tagged with a found_on array listing which competitors it links to. The composite tool, and the decision I'm unsure about Most API-wrapper MCP servers are pure 1:1 mappings. I added one opinionated composite tool, gap_outreach_targets , because the raw gap output isn't the thing you actually want — it's the raw materia

I once submitted an essay with three citations that I hadn't personally verified. The AI had suggested them, and they sounded right. None of them existed. That's not a quirk or a bug — it's exactly how LLMs work. And once you understand why, a technique called RAG starts to make a lot of sense. AI assistants are remarkably good at sounding right. The model isn't lying — it's doing its best with what it knows. The problem is that what it knows has limits, and it doesn't always know where those limits are. Ask one about a recent event, a niche regulation, or anything from a source it's never seen — and it fills the gap anyway. Confidently. That's the gap RAG was built to close. Once you understand how it works, you'll have a much clearer picture of why some AI tools are genuinely reliable and others are just very convincing guessers. Here's what's actually going on. First, What's the Problem? Large language models (LLMs)—the technology powering AI assistants like ChatGPT and Claude—are trained on vast amounts of data from across the internet. That training gives them a remarkable ability to reason, summarize, and generate content. But it also comes with some real limitations: They have a knowledge cutoff. An LLM trained last year doesn't know what happened last month. They can hallucinate. When they don't know something, they don't say "I don't know"—they generate a confident-sounding answer anyway. Wrong facts, fake statistics, invented sources. All delivered with a straight face. They don't know your specific sources. Think of a software engineer asking an AI assistant about their company's internal API documentation, deployment runbooks, or architecture decisions. None of that is in the training data. The model has never seen it — and it will still try to answer. The model isn't lying — it's generating the most plausible answer it can. It just has no way to know when it's wrong. So, what do you do when you need an AI that's accurate, current, and knows your specifi

"Settlement layer for the agent economy" has become a phrase used by more than one architecture this month. Some of those architectures are venues. Some are payment rails. Some are identity stacks. One is what we build — a trust-minimized atomic settlement primitive — and we keep getting asked how it sits next to the others. So this is a short, opinionated week-in-review. Seven moves that landed recently in agent-commerce infrastructure, what each one actually is, and where atomic settlement fits underneath. No leaderboard. No "who wins." Just the layers. 1. Coinbase's Base MCP (May 26) Coinbase shipped a Base MCP server: ChatGPT, Claude, and Cursor can connect directly to Base wallets, with built-in Uniswap and Morpho integrations. Two-line install, agent-driven swaps, on-chain account control from inside the model client. What it is: a wallet-side, exchange-aligned MCP. The agent drives a wallet on Base; Coinbase has thought hard about UX, key handling, and AI client ergonomics. If your agent is already operating inside the Coinbase + Base orbit, this is a meaningful drop in friction. What it isn't: a cross-chain settlement layer. It is one L2, deeply integrated. The agent's trust assumptions are "Coinbase wallet infrastructure + Base + the DEXes that Base MCP integrates with." That is a reasonable trust budget for a lot of trades; it is not the budget for an agent that wants to trade BTC for ETH without picking a venue. 2. The x402 Foundation (Coinbase + Cloudflare) Coinbase and Cloudflare announced an x402 Foundation this month. x402 is the HTTP-based agent payment standard — a 402 response code carrying payment metadata, designed so an agent can pay for an API call inline. Cloudflare joining elevates this from a Coinbase-led initiative to a multi-party standards effort. Cloudflare also runs an enormous slice of the internet's edge; if x402 becomes a default payment rail, it will be visible from the edge first. What it is: a payment-handshake protocol. How an ag

Every token your agent spends on file I/O is wasted reasoning capacity. I was building a document processing agent — the kind that reads incoming research reports, extracts key findings, and produces executive briefings. Nothing exotic. The kind of workflow thousands of teams are automating right now. The PDF I was testing with was 2MB. Dense text. A typical industry research report. When I measured the token cost of processing it inline, the number was 97,354 input tokens — just to get the text into Claude's context. At claude-sonnet-4-6 pricing, that's $0.29 per document. For a pipeline that processes 500 reports a month, you're looking at $150/month before your agent writes a single word of output. That's the problem nobody talks about in the AI agent space. Everyone optimises prompt engineering and output tokens. The silent cost is input: the files, the content, the raw data you're shoving into context before the agent can do anything useful. How the token count explodes When you pass a document to an agent inline, one of two things happens: Option A — Base64 encoding. You read the binary file, encode it, embed it in the prompt. A 2MB PDF in base64 is ~2.7MB of text. At roughly 3.5 characters per token, that's ~770,000 tokens before your agent has read a single word. This is catastrophic. Don't do this. Option B — Text extraction. You extract the raw text content first (via pdftotext , PyMuPDF, or equivalent), then pass the text to the agent. Better — but a 2MB PDF with dense content still yields ~97,000 tokens of extracted text. You've paid for every word, every header, every footnote. Either way, the document content dominates your context window, crowds out your system prompt, and you're burning money on file I/O instead of reasoning. The alternative: specialist services via MCP Model Context Protocol (MCP) is Anthropic's open standard for connecting AI agents to external tools and services. The key insight is simple: your agent doesn't need to contain the co

This is a submission for the Hermes Agent Challenge : Build With Hermes Agent What I Built Real brewing knowledge lives in human experience — in roaster guides, in community notes, in what a barista learned from last Tuesday's pour. It doesn't accumulate anywhere. Every brew is forgotten. Ask any AI and you get statistical averages: 93°C, 1:16 ratio, four minutes. Technically defensible. Practically generic. Worse still for rare origins where training data is thin. Demo For coffee drinkers Visit brew-guide-production.up.railway.app . No account. No setup. No AI client required. Pick your coffee origin, roast level, and brew method. What comes back isn't a generic recipe — it's community consensus: the grind, temperature, ratio, and brew time that real people have logged and rated for that origin, plus step-by-step technique guidance (bloom timing, pour stages, agitation style). If data is sparse for your origin, the confidence tier says so honestly and falls back to method defaults rather than making something up. This is for the person who just picked up a bag of Kenyan peaberry and wants to know how to do it justice. It works for anyone who cares about their cup — no technical knowledge required. For developers and AI clients Connect to any MCP-capable client in one line: https://brew-guide-production.up.railway.app/mcp Ask your AI: "recommend a pour over for Ethiopian light roast." What comes back is a traceable community consensus object: brew parameters, a confidence tier (high/medium/low), the source brews that contributed, and method-specific technique guidance. You can see where the knowledge came from and how certain the system is — a fundamentally different epistemic object from an AI-generated recipe. Code GitHub: yuens1002/brew-guide Five MCP tools — get_brewing_methods , recommend , log_brew , search_brews , compare_brew — over Streamable HTTP transport. Public, no auth required. My Tech Stack Layer Technology HTTP Hono 4 + @hono/node-server MCP @modelc

Google I/O 2026: MCP Is Now Infrastructure Google I/O used to be about new models. This year it was about what those models do - and how they connect to everything else. MCP was everywhere. Not as a novelty. Not as an experiment. As the assumed plumbing. Here's what actually shipped. Gemini Spark Will Run on MCP for Third-Party Tools The headline agent at I/O 2026 was Gemini Spark - a 24/7 AI agent that runs on cloud VMs, works while your devices are off, and handles long-running tasks across Gmail, Docs, and Calendar. Spark integrates with Google Workspace apps first, then expands to third-party tools via MCP over the summer. That's the part worth sitting with. Google built its flagship consumer agent and then said: for everything outside our walls, we'll use the open protocol. A year ago, MCP was a specification from Anthropic. Today, Google built its flagship consumer AI agent on it. Cursor, Copilot, Windsurf, Mistral, Grok - they all support it too. When the company that runs Search, Gmail, Android, and Chrome commits to MCP as the integration layer for its flagship product, the protocol debate is effectively over. Managed Agents Get MCP Servers by Default Google also launched Managed Agents through the Gemini API - a setup where a single API call provisions a remote Linux environment with its own isolated sandbox. Each agent gets its own ephemeral sandbox provisioned with skills, Model Context Protocol (MCP) servers, and server-side tools. Full integration with A2A and Agent Platform governance and security are coming soon. Managed Agents are powered by the Antigravity agent and built on Gemini 3.5 Flash. Developers can define custom agents through versionable markdown files such as AGENTS.md and SKILL.md, rather than building complex orchestration layers from scratch. This is Google offering hosted execution, sandboxing, state handling, and MCP tool access as a bundled service. The enterprise pitch is operational abstraction - you define the agent, Google runs

Cloudflare recently added support for Claude Managed Agents, allowing developers to run and manage Claude agents within Cloudflare. Developers can connect agents to private systems, choose their runtime environment, and monitor agent activity using Cloudflare services. By Renato Losio

"Settlement layer for the agent economy" is suddenly a crowded sentence. In the last few weeks, two very different things have started competing for it. OKX Agent Payments Protocol (APP) announced in May 2026 — backed by AWS, Alibaba Cloud, Uniswap, Paxos, QuickNode, with ecosystem support from Base, Ethereum Foundation, Solana, Sui, Aptos, Optimism — describes itself as the settlement layer where AI agents pay each other. AEON's $8M pre-seed (May 20, YZi Labs) described itself the same way. There is a list now, and it is getting longer. We build one of the things on this list, and we think the framing is wrong. These are not rivals jostling for the same slot. They are two different layers, and they answer two different versions of the same question: what does an agent need to settle a trade? This piece walks through the two layers, where each is the right answer, and why an honest comparison gets you further than picking sides. What APP and other custodial-venue protocols actually do A useful way to read OKX APP — and similar moves coming from the larger exchanges — is to treat them as a venue layer . An exchange already holds inventory across many chains. It already has the risk engines, the liquidity, the legal arrangements with banks and partners. Adding an agent-facing API on top of that machinery is a relatively short walk. What an agent gets, in exchange, is breadth. Many chains, many assets, batched and netted settlement against deep internal books, and fast execution because the exchange is just moving entries in its own ledger. For an agent whose job is "find a price somewhere and execute now," that is a powerful primitive. What the agent gives up, in exchange, is a counterparty. At any moment between deposit and withdrawal, the agent's balance is the venue's promise to pay. That promise is normally good. The agent has no way to verify, from inside its own logic, that it still is. This is not a criticism of OKX APP. It is the structural shape of any venue-

This is a submission for the GitHub Finish-Up-A-Thon Challenge What I Built Last...

A while ago I tried to build a local coding assistant. I downloaded Qwen3, fired it up on my MacBook with 16GB of RAM, and within a day realized the output quality was nowhere close to Claude or GPT-5. The model could fit . It just couldn't compete . So I changed the question. If I can't make the model smarter on my hardware, can I make what I feed it smarter? Where the tokens actually go I started watching where my Claude / Cursor / Copilot sessions actually spent their tokens. The surprise: most of it wasn't reasoning. It was lookup . Every fresh chat about my company's database re-discovered the same things: What does status = 3 mean? (cancelled) How does orders join to users ? ( orders.user_id → users.id ) What's that cryptic JobStatus enum? (a dozen integer codes nobody remembers) The model figured it out, the session ended, and tomorrow it figured it out again . Same tokens, same latency, every single time. The expensive part of working with an AI wasn't the thinking — it was re-teaching it things it had already learned yesterday. There's a lot of attention right now on trimming AI output tokens (talk like a caveman, strip the pleasantries, etc.). But in my workflow the bigger leak was on the input side: paying full token cost every session to re-establish context that never changed. "Memory" isn't a feature, it's an architecture question AI clients are starting to bolt on "memory" features. But they're proprietary, opaque, and locked to one tool. Claude's memory doesn't help Cursor. Cursor's doesn't help Copilot. You can't inspect it, you can't share it with a teammate, and you can't diff it. What I actually wanted was an explicit, inspectable, shareable context layer that any AI client could read deterministically — same answer every time, same file my team could hand off. I picked the highest re-learn cost in my world to start with: SQL databases. Enter amnesic amnesic is an open-source MCP server that gives any AI client persistent semantic memory of your