Open Source Radar — November 10: Home Assistant Gets an MCP
Today's board is about plugging an agent into something you already run: a house, a memory, a portfolio. All three are MIT-licensed, and all three moved from "nice demo" to "installed" this week.
ha-mcp (Python, about 6,000 stars, MIT) lets Claude, ChatGPT or any other MCP client drive Home Assistant in plain language — 87 tools covering device control, state queries, automations, scripts and dashboard cards, plus execution-trace analysis for when an automation silently stops firing. The reason it is on the daily board today is packaging rather than features: since version 9.0.0 on October 28 it installs as a Home Assistant integration itself, through HACS or the Marketplace on Home Assistant 2026.11, and runs the whole server in-process on OS, Supervised, Container and Core installs alike — no sidecar container, no separate process, and remote access through a Home Assistant webhook whether that is Nabu Casa or your own reverse proxy. The project is upfront about the trade: setup asks which administrator account the server acts on, because an agent that can write automations can also break them, and 19 open issues against 6,000 stars is a healthy ratio rather than a warning. The useful version of this is boring — asking for an automation that turns on the porch light at sunset and getting working YAML back, instead of a paragraph explaining what a trigger is.
gbrain (TypeScript, about 32,000 stars, MIT) is a memory store you own, built for the agent you already run rather than as a new agent: it records explicit facts with their sources, allows corrections and withdrawal, and serves the same memory to Claude Code, Codex, Cursor or a chat assistant over MCP. The default is deliberately unglamorous — keyword retrieval with no embedding service and no API key — with semantic search, synthesis and background enrichment as opt-ins, which is the right order for anyone who has watched a memory layer invent a preference they never stated. It is Garry Tan's repo, seven months old, and the star curve says people want their context to survive a model switch; the 809 open issues say it is still early. Reach for it when your agent's memory lives inside one vendor's account and you want it portable — with one caveat, the hosted version at gbrain.io is a paid product while the self-hosted repo is free.
daily_stock_analysis (Python, 69,411 stars, MIT) is the retail-trader version of the same idea: an LLM watchlist analyst across A-shares, Hong Kong, US, Japan, Korea, Taiwan and ETFs that pulls quotes and news from multiple sources, scores each holding, and pushes a decision dashboard — buy and sell levels, risk alerts, catalysts — to WeCom, Feishu, Telegram, Discord, Slack or email on a scheduled run that costs nothing on GitHub Actions. When we looked at it in August it had about 61,600 stars; it has added roughly 8,000 since, and the number worth staring at is the fork count: 56,566 forks against 69,411 stars. That ratio is not admiration, it is deployment — forking is how you get a private copy running on a schedule with your own secrets and your own watchlist. It ships 15 built-in strategies and a multi-turn question mode for follow-ups, defaults to free data sources that are rate-limited and occasionally break, and says plainly that it is analysis, not investment advice.
Worth watching this week: whether a Marketplace listing turns MCP servers into a normal app category, and whether gbrain's issue count falls as fast as its star count rose.
Would you give an agent admin rights over your house before you give it your brokerage watchlist? Write to us via our contact page.
