Beat: publicly available, usable-now developer artifacts — tools/skills/MCPs we can ADOPT instead of build. Not news (ai-watch), not model releases (model-watch), not research (Sol). Run by code-eth, weekly. Reads: Apps/app-architecture_roadmap.md wishlist · eth-memory/shared/intelligence-board.md.

This is the first run, so the bar is "what would have changed a build decision if we'd known." Four of these map straight onto wishlist items we have queued to build by hand. The headline: the house-automation spine and the Communication Hub are both off-the-shelf, MIT-licensed, and actively maintained — we should not hand-roll either.

ADOPT-CANDIDATEs (verified: real repo, real commits, fits our stack)

ha-mcp — Home Assistant MCP server → unlocks the whole "house automation" line

• Repo: https://github.com/homeassistant-ai/ha-mcp · MIT · ⭐3.3k · v7.7.0 2026-06-10 (1,623 commits — very alive)
• WHAT: ~83 tools letting Claude read/control Home Assistant by natural language — devices, automations, dashboards, history, backups. Runs locally (Docker or pip on a Pi), data stays on the LAN, no telemetry.
• WHY US: This is the missing spine for the telescope mount, the bathhouse, and the 4 chassis. The roadmap has no "house automation" app yet because we assumed we'd build the bridge ourselves — we don't have to. Home Assistant is the device layer; ha-mcp is the Claude↔devices bridge, already at v7.7. Eth-on-Pi already speaks MCP. Adopter: whoever stands up the house layer (Eth-Pi embodiment) — install HA on a Pi, point ha-mcp at it, expose only the entities we want Claude to touch.
• Verdict: ADOPT-CANDIDATE. Highest-leverage find this week. Risk is scope (83 tools = wide blast radius); mitigate by exposing a narrow entity set first.

OpenClaw — self-hosted multi-channel assistant gateway → replaces the Communication Hub build (roadmap Priority 2)

• Repo: https://github.com/openclaw/openclaw · MIT · ⭐ very large but SEO-inflated across sources (real repo, 68k+ credibly; 180k–340k claimed — treat the big numbers with suspicion)
• WHAT: Runs on your own box; connects 20+ chat channels (Telegram, Signal, WhatsApp, Discord, Matrix, iMessage, WebChat…) to local or remote model backends. Orchestration/routing layer; skills + context stored locally.
• WHY US: The roadmap's Communication Hub ("push to phone, two-way, replaces the Telegram bot") is exactly this, and it's the thing we keep deferring as "medium complexity — needs service worker + push." OpenClaw gives two-way on channels Zaina already uses without us writing the push plumbing. Adopter: Eth-Code, as the Telegram-bot replacement path.
• Verdict: ADOPT-CANDIDATE, with a caveat. It's a gateway, not a voice — it should sit in front of our existing voices, not replace the bespoke Pi/voice architecture. Evaluate whether it cooperates with eth-state/our MCP servers or wants to own the loop. Worth a half-day spike before committing.

last30days — multi-platform research skill → strengthens our own watch lanes (ai-watch, tooling-watch, Sol)

• Repo: https://github.com/mvanhorn/last30days-skill · MIT · ⭐40.1k · v3.3.0 2026-05-17 (623 commits)
• WHAT: A Claude Code skill (/last30days <topic>) that fans out across Reddit, X, HN, YouTube, Polymarket, Bluesky, TikTok + web, engagement-weights and de-dups, and synthesizes a cited brief. Zero-config for the free sources (Reddit/HN/Polymarket/GitHub/X-via-browser-login); Brave key (2k free/mo) and yt-dlp optional.
• WHY US: This is the fan-out-and-synthesize pattern every watcher (this one included) hand-rolls each run. It would not replace the analysis/angle the beats contract demands, but it replaces the search-and-collect half — especially the social/community surfaces (Reddit/X) we currently miss. Directly serves Zaina's "my Instagram is full of tools we haven't thought of" motivation: it reads the social layer we don't. Adopter: all watch lanes; install as a shared skill, run before the human-angle pass.
• Verdict: ADOPT-CANDIDATE. Low risk (read-only, local), high recurring value. Install and trial on next week's tooling-watch.

Bumblebee — read-only supply-chain scanner (incl. MCP configs) → serves our permissions/risk posture

• Repo: https://github.com/perplexityai/bumblebee · Apache-2.0 · ⭐~2.9k · v0.1.1 May 2026 (single static Go binary, zero non-stdlib deps)
• WHAT: Read-only scan of on-disk lockfiles, package-manager metadata, IDE/browser extensions, and MCP JSON configs; emits structured NDJSON and flags matches against a known-compromise catalog. Three profiles (baseline/project/deep).
• WHY US: We run several custom MCP servers, install community skills/tools (this very report recommends four more), and keep plaintext creds by choice. A read-only audit of exactly that surface — MCP configs + extensions + deps — is the cheapest possible guardrail. Single binary, nothing to maintain. Adopter: Eth-Code, as an occasional hygiene scan before/after adopting new tools.
• Verdict: ADOPT-CANDIDATE. Lowest-effort item here; pure downside-protection given we're actively pulling in third-party artifacts.

WATCH (real, relevant, not yet a decision)

• Agent-memory MCPs — Hindsight (vectorize.io), OpenMemory (Mem0), AgentMemory (rohitg00). Persistent retain/recall/reflect memory servers, MCP-compatible, local-first. Why only WATCH: we already run a bespoke memory + shared-state system (eth-memory worker, sol-memory-worker, eth-state) with cross-voice semantics these don't have. Switching is high-risk for marginal gain. Worth reading their reflect/auto-update designs as prior art for our own, not adopting. — https://github.com/rohitg00/agentmemory · https://mem0.ai
• Read-only finance MCPs (10-K/filings/prices/portfolio-review). Could feed the Portfolio Dashboard (roadmap Priority 2, "high complexity, needs IB pipeline") with fundamentals/prices without us building feeds. WATCH which ones are MIT and broker-agnostic. — https://vibehackers.io/mcp/category/finance-payments
• Codex plugins (OpenAI, shipped 2026-06-02). Bundle skills + app integrations + MCP configs into one installable for Codex. We run Sol-on-Codex with hand-edited ~/.codex/config.toml; this could package Sol's MCP/tool set cleanly. WATCH for whether it simplifies the sol-memory-server stdio proxy setup.
• Brave Search MCP (2k free queries/mo) / Exa. Agent-native search backends. last30days already uses Brave; if we adopt it, the same key backs our watchers' web search instead of ad-hoc. Cheap WATCH.

SKIP-but-noted

• Alpaca MCP — live trade execution ("buy 10 shares of AAPL" → real order). SKIP: violates the house rule (never execute trades/move money) and we're on IB, not Alpaca. Noted only so a future watcher doesn't re-surface it as new.
• Langflow / general agent-builder frameworks (LangChain-based). SKIP: heavyweight, framework-lock-in, against the roadmap philosophy ("no frameworks = no dependencies"). Not for us.

Bottom line

Two builds we had queued — house automation (no app yet) and the Communication Hub (deferred) — both have mature, MIT, actively-maintained off-the-shelf answers (ha-mcp, OpenClaw). One skill (last30days) upgrades the watch lanes themselves; one binary (Bumblebee) is free insurance for the third-party tools we keep adding. The house with the telescope and the bathhouse gets its nervous system from ha-mcp — that's the find of the week.

Next run: trial last30days on the watch itself; check whether OpenClaw cooperates with eth-state or wants the loop.