harrylabsj

Eatwise (外卖营养教练) — local-first Chinese dietary coach: deterministic meal nutrition reviews, personal daily energy/nutrient targets, and consented on-device profile and meal history through 9 MCP tools over a bundled open food database (TFND + USDA FNDDS). No network, no accounts, no payments; profile schema rejects name/phone/ID/medical fields by design. Engine pinned to an immutable git commit (cbc08f6) over npx stdio; Portable Agent Plugins v1 package (mcp.json + skills/), needs Node.js 22.5+ on PATH. Disclosure — the engine's MCP tool descriptions and error messages are Chinese-only while the skills are English-first.

Payer-side freight reconciliation: import carrier bills and trip ledgers into a deterministic local core (SQLite, WAL, no cloud), re-calculate against contracted rates and produce evidence-linked difference lists for human review — 14 MCP tools over uvx stdio (engine pinned to the immutable commit e8cd259c, release v0.4.1), with 3 bilingual workflow skills bundled. Portable Agent Plugins v1 package (mcp.json + skills/); needs uv on PATH. Never approves payments, sends messages or writes back to ERP/TMS; final confirmations route through a local admin page the model cannot reach. Disclosure — the optional freight-admin page listens on 127.0.0.1 and is where a human confirms mappings, rules, matches and exports; nothing leaves the machine.

Wrong-question diagnosis and study-action loop for parents and students — deterministic local core (SQLite, no cloud), host-model recognition gated by parent confirmation. Portable Agent Plugins v1 package (mcp.json + skills/); good-student MCP server pinned to immutable commit 2469cbb over uvx, with the good-student workflow skill bundled. Needs Python 3.10+ via uv. No credentials required; all data stays on-device and nothing is uploaded.

Local-first personal health records: encrypted on-device Vault (AES-GCM) for vitals, medications, activities, emotions, sleep, encounters, diagnosis mentions, medication plans and reminders, served through a Keychain-paired local daemon with 24 scoped MCP tools. Zero-config first run: the plugin bootstraps Vault + daemon + agent pairing automatically under PLUGIN_DATA. Engine pinned to an immutable PyPI release (healthcare-mcp==0.2.1) over uvx stdio; Portable Agent Plugins v1 package (mcp.json + skills/), macOS only, Python 3.11+. Zero network by default; the optional LLM report-import endpoint stays off unless the user configures it. Approval-gated deletes, tamper-evident audit chain. Disclosure — personal health record-keeping, not medical advice (the skill forbids diagnoses and dosage changes); first run has uvx fetch healthcare-mcp==0.2.1 from PyPI; the optional report-import path posts report images to Alibaba DashScope (qwen-vl-max) unless HEALTHCARE_LLM_ENDPOINT points elsewhere, and only when HEALTHCARE_LLM_API_KEY is set.

Source suppliers and negotiate purchases through the Kiwi Buyer MCP server — an open-source A2A commerce runtime (pinned @harrylabsj/kiwi@0.12.5 over npx stdio), with the kiwi-buyer sourcing workflow skill bundled. Sourcing is dual-source: Kiwi Network results come from kiwi_search, and internet e-commerce listings are reported through the host's own web tools when the session has them (a host capability, not a plugin guarantee). Portable Agent Plugins v1 package (mcp.json + skills/); needs Node.js/npm on PATH. No credentials required; Kiwi never handles payments.

Local-first WeRead (微信读书) reading companion: syncs bookshelf, reading stats and notes to the local machine and turns them into private reading cards and weekly/monthly reviews through 8 MCP tools, with a built-in fictional demo mode for previews. Engine pinned to an immutable git commit (e5f6673) over npx stdio; Portable Agent Plugins v1 package (mcp.json + skills/), needs Node.js 20+ on PATH. All data stays in PLUGIN_DATA; never posts to WeRead — cards and reviews are local files saved only after user confirmation. Disclosure — live mode sends your WEREAD_API_KEY as a Bearer token plus book ids and paging cursors to i.weread.qq.com (Tencent's WeRead agent gateway) and nothing else; the engine's MCP tool descriptions and error messages are Chinese-only while the skills are English-first; the engine's npm dependencies resolve as caret ranges at first launch.

Local-first WeChat relationship memory assistant for macOS: captures only what is visible in the WeChat window via system OCR/Accessibility, stores everything in a local SQLite database (0600/0700 permissions, no cloud), and serves contact briefs, follow-up planning, outreach drafts and weekly relationship reports through 32 MCP tools. Engine pinned to the immutable PyPI release wechat-social-assistant==1.5.1 over uvx stdio; Portable Agent Plugins v1 package (mcp.json + skills/), needs uv on PATH. All data confined to PLUGIN_DATA via WSA_ALLOWED_ROOT + path enforcement. Never reads WeChat databases, never injects into WeChat, never sends messages — outreach is drafts for human review only. Disclosure — reads WeChat only by screenshotting/OCR-ing and walking the macOS Accessibility tree of the frontmost window (not WeChat's databases or a bot API), compiles bundled Swift helpers with swiftc into ~/Library/Caches/wechat-social-assistant/bin on first use, and needs Screen Recording (optionally Accessibility) permission for the host terminal; no network egress, never sends messages.