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 first: watchlist.md · Apps/app-architecture_roadmap.md · eth-memory/shared/intelligence-board.md.

Honest framing: this was a watchlist-driven week. The two named entries resolved cleanly (one ADOPT, one hard SKIP), and the needs-specific sweep turned up one genuinely strong embodiment find. The generic GitHub/MCP sweep was mostly aggregator listicles ("337 skills!", "425 plugins!") with no single new artifact that would change a build decision — so I'm not padding. One real generic candidate this week, plus the watchlist resolutions.

WATCHLIST RESOLUTIONS (these outrank everything)

Obsidian + AI ecosystem → ADOPT-CANDIDATE (Zaina, 06-12)

• Pick: markuspfundstein/mcp-obsidian — https://github.com/markuspfundstein/mcp-obsidian · MIT · ⭐3.9k · 71 commits, 60 open issues / 30 PRs (alive, community-active). Node alternative exists: cyanheads/obsidian-mcp-server.
• WHAT: MCP server giving Claude 7 tools over an Obsidian vault — list_files_in_vault, list_files_in_dir, get_file_contents, search (full-text/fuzzy), patch_content (insert relative to headings), append_content, delete_file. Talks to the vault through the Obsidian Local REST API community plugin; runs locally, notes never leave the device.
• WHY US: We have markdown memory everywhere — eth-memory, Sol rooms, Cas spine, the Architecture system-of-record, this very report. The thing this unlocks is structured search + patch over a vault without us writing a retrieval layer. If/when any of our markdown homebases moves into an Obsidian vault (Pinehaven used exactly this pattern as their homebase — board 06-12), this is the off-the-shelf bridge. Even today it's a clean way to let Claude Desktop read/patch a local notes vault by natural language. Local + MIT + no telemetry = fits permissions/risk posture.
• Caveat: requires the Obsidian Local REST API plugin running (one local dependency). delete_file is a write tool — expose read/search first, gate deletes. Verdict: ADOPT-CANDIDATE — trial as a Claude Desktop connector against a throwaway vault before pointing it at real memory.

"ClawCode" → SKIP (with the legitimate need flagged) (Zaina, 06-12)

• What it actually is: the family-known "ClawCode" is the cluster of leak-derived Claude Code rewrites that exploded after Anthropic's Claude Code source was exposed 2026-03-31. Lead: Sigrid Jin (WSJ-profiled 03-21 as a top Claude Code user, ~25B tokens) — ported the harness to Python "before sunrise," now a Rust rewrite "in progress." Surfaces: claw-code.io, claw-code.codes, github.com/hinet/claude-code (claims 50k⭐ in 2h), github.com/Kuberwastaken/claurst. Several frame themselves "clean-room… beyond any DMCA reach."
• What it unlocks (the real question): an open, model-flexible agent harness — i.e. run a Claude-Code-style coding loop against non-Anthropic backends (DeepSeek/local), which is literally our Plan B (roadmap "Platform Independence via DeepSeek V4"). That need is real.
• Why SKIP anyway: provenance is leaked proprietary Anthropic source, however much the wrappers say "clean-room." We are a house that builds publicly ON Anthropic (EthBuilds, Cowork, Claude Code is Eth's daily substrate). Adopting a leak-derived harness carries legal/ToS, security (running an unofficial harness with our creds + custom MCP servers), and reputational risk to the EthBuilds brand that vastly outweighs the convenience. No license on the credible repos = unlicensed. Verdict: SKIP.
• The clean path for the same need: when Plan B gets real, reach for a legitimately-licensed model-agnostic agent CLI (e.g. genuinely open harnesses with their own clean lineage) or build the loop on our own terms — not the leak. Logged so a future watcher doesn't re-surface ClawCode as a fresh "look, model-flexible Claude Code!" find.

ADOPT-CANDIDATE (generic — 1 this week)

xiaozhi-esp32 — voice AI on ESP32, MCP-native → the off-the-shelf voice/embodiment hardware layer for the chassis

• Repo: https://github.com/78/xiaozhi-esp32 · MIT · ⭐27.4k · v2.2.6 2026-04-19 (64 releases, 924 commits — very alive)
• WHAT: Open-source AI voice chatbot firmware for ESP32 microcontrollers. Offline wake-word, streaming ASR → LLM → TTS, speaker recognition. Runs on 70+ cheap boards (ESP32-C3/S3/P4; M5Stack, Waveshare, LILYGO). MCP-native — uses Model Context Protocol for device control and to extend the model's capabilities. Backend is configurable: defaults to xiaozhi.me but supports custom servers and DeepSeek/Qwen. Docs in EN/中文/日本語.
• WHY US: This is a real answer to two lanes at once. (1) Embodiment / the 4 chassis — instead of hand-rolling a wake-word→STT→TTS loop on each chassis, this is a maintained, $5–15-board voice frontend that already speaks MCP, so it sits in front of our architecture rather than owning the loop (the cooperation test OpenClaw still has to pass — this one passes by design). (2) The voice lane's pain — our current Pi voice path is a Jabra-mic-dependent loop with the recurring capture-volume timeout (eth_voice_mic_timeout); a dedicated voice MCU offloads wake/ASR to hardware built for it. (3) Plan B alignment — native DeepSeek/Qwen backend support means a chassis voice that doesn't depend on the Anthropic tier that just proved it can be export-frozen (ai-watch Fable saga).
• Caveat: it's Chinese-ecosystem-first (default cloud is xiaozhi.me — point it at our own backend, don't use the default for anything private); v2 firmware is a fresh flash (v1-incompatible). Hardware project — needs someone to flash a board, not a pip install.
• Verdict: ADOPT-CANDIDATE. Buy one ESP32-S3 board, flash v2.2.6, point it at a local/DeepSeek backend, expose a narrow MCP tool set — cheapest possible spike of "what does a real voice chassis feel like" before we commit to building chassis voice by hand. Adopter: Eth-Pi embodiment.

WATCH (real, relevant, not yet a decision)

• Robotics MCP servers, broadly — ChatForest's roundup (chatforest.com/reviews/robotics-mcp-servers) catalogs MCP servers for ROS, ESP32, robot arms, drones, physics sim, incl. DimOS (agentic robotics OS) alongside ha-mcp. As the chassis/telescope-mount line moves from idea to build, this is the reference shelf — most of the "make Claude move a physical thing" plumbing already exists as MCP. WATCH; revisit when a specific chassis actuator is chosen.
• ha-mcp tool-count bump — this week's roundups now cite 86 tools (was 83 at v7.7.0 last week). Minor, but confirms the house-automation spine from edition 1 is still shipping fast. Still the ADOPT for the house layer; no change to the 06-12 verdict.
• OpenClaw cooperation check — STILL PENDING (carried from 06-12). Now past ~210k⭐ and named the fastest-growing OSS project of 2026. The half-day spike — does it cooperate with eth-state / our MCP servers or want to own the loop — hasn't been run. Note: xiaozhi-esp32 answers the voice-frontend slice of the Communication-Hub need MCP-natively, which may lower OpenClaw's urgency; OpenClaw is still the better fit for the multi-channel chat gateway (Telegram/Signal/WhatsApp) slice.

PENDING FROM LAST WEEK (honest carry-forward)

• last30days trial — NOT DONE. Edition 1 committed to trialing the /last30days skill on this run to upgrade the search-and-collect half (esp. the Reddit/X/IG-adjacent social layer that is exactly Zaina's "my Instagram is full of tools" gap). It was not installed/trialed this week — this sweep was still hand-rolled web search. Action stands: install last30days as a shared skill and run it before next week's human-angle pass. This is the single highest-leverage process upgrade for this beat and it keeps slipping.

SKIP-but-noted

• The "337 skills / 425 plugins / 2,810 skills" mega-marketplaces (alirezarezvani/claude-skills, jeremylongshore/clau…REDACTED, tonsofskills.com, etc.) — SKIP as bulk. These are aggregators, not artifacts; pulling a 300-skill bundle is the opposite of our "every dependency is a point of failure" philosophy and exactly the surface Bumblebee (edition 1) exists to scan. If a specific skill from one of them ever matters, we adopt that one skill, audited — never the bundle.

Bottom line

Watchlist-driven week, honestly thin on generic finds. Two clean resolutions: Obsidian MCP is an ADOPT (off-the-shelf vault↔Claude bridge for the markdown memory we have everywhere); ClawCode is a SKIP (leak-derived — the legitimate "model-flexible coding harness" need is real and points at Plan B, but not via stolen source, not for a house that builds publicly on Anthropic). The one real generic find, xiaozhi-esp32, is genuinely good: a maintained, MIT, MCP-native, DeepSeek-capable voice hardware layer that could be the chassis voice frontend we were about to hand-build — and it sidesteps both the Jabra-mic timeout and the export-frozen-tier risk in one move. Two process debts remain open and named: trial last30days, spike OpenClaw cooperation.

Next run: actually install + run last30days before the sweep; if any chassis hardware is chosen, pull the Robotics-MCP shelf forward to ADOPT.