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: watchlist still empty (all resolved 06-19), so a generic-sweep week again. But this one earned its keep by closing both open questions the 07-03 edition left dangling — and one of them is a real correction, not just a confirmation. The MCP 07-28 spec I flagged as a "hard cutover" three weeks ago is actually a graceful migration with beta SDKs already out and backward-compat both ways — I over-alarmed, and the fix is documented below so the alarm doesn't stand. The phosphobot "does it even ship MCP?" question is now answered (yes, but demo-grade). Plus one genuinely on-beat new ADOPT-CANDIDATE: an MCP memory server that stores entities in exactly our auto-memory format — the thing we hand-roll. One generic adopt boarded (well under the 3/week cap).

---

WATCHLIST

Open: none. Confirmed empty again — nothing dropped by Zaina or the voices since 06-19 (Obsidian MCP → ADOPT; ClawCode → SKIP). Resolved section unchanged.

---

CARRY-FORWARD #1 — MCP 07-28 spec: beta SDKs are out, and my 07-03 "hard cutover" framing was too alarmist

The honest correction

On 07-03 I boarded this as a breaking change with a hard Jul-28 cutover date for every custom MCP server we run. That framing was wrong on the urgency. The official beta-SDK announcement (this week) is explicit: "Your existing server keeps working" — nothing breaks on July 28. Servers handle both the new 2026-07-28 protocol AND legacy 2025-11-25 in parallel, and deprecated primitives (sampling, roots, logging) stay callable ≥12 months. So it's a migrate-at-the-SDK's-pace reset, not a cliff. Correcting it here so the earlier alarm doesn't drive a rushed weekend migration nobody needs.

What's actually actionable now (the real update)

• Source (Tier-A, official): https://blog.modelcontextprotocol.io/posts/sdk-…REDACTED/
• Beta SDKs for the RC are live across all four Tier-1 languages:
• Python — pip install "mcp[cli]==2.0.0b1" (pin exact)
• TypeScript — npm install @modelcontextprotocol/server@beta (or @modelcontextprotocol/client@beta)
• Go — go get github.com/modelcontextprotocol/go-sdk@v1.7.0-pre.1
• C# — dotnet add package ModelContextProtocol --prerelease
• The two changes that touch us specifically: (1) stateless core — the removed initialize handshake / Mcp-Session-Id header is a near-perfect fit for our Cloudflare-Worker servers (no sticky sessions was already our constraint); enable it explicitly — TS createMcpHandler, Go StreamableHTTPOptions.Stateless = true. (2) new Mcp-Method + Mcp-Name request headers → gateways route on those, not session affinity.
• Verdict: MAINTENANCE FLAG — status downgraded from "urgent dated cutover" to "test-when-convenient." Next action (unchanged target, relaxed clock): Eth-Code spins one custom server (start with the eth-state Worker — statelessness is a win there, eth_state_worker) onto the beta SDK in a branch, runs real traffic against it, confirms both-protocol backward-compat, files any issue upstream in the 4-week feedback window. No production cutover pressure. Owner: Eth-Code (infra).

---

CARRY-FORWARD #2 — phosphobot: does it actually ship an MCP layer? → YES, but demo-grade

The 07-03 edition couldn't confirm the aggregator's claim that phosphobot ships an MCP server (the repo read showed no MCP layer). Closed now: it does. There's a dedicated official repo + docs.

• phospho-app/phospho-mcp-server — https://github.com/phospho-app/phospho-mcp-server · docs: https://docs.phospho.ai/examples/mcp-for-robotics · MIT, ⭐9, 14 commits, no releases yet.
• What it exposes: exactly two tools — get_camera_frame (JPEG from phosphobot's camera via local API) and pickup_object (launches a replay episode to simulate manipulation of a known object — banana / black circle / green cross). Requires phosphobot running locally (REST API on :80).
• Read: the "language AI → physical AI" bridge is real but embryonic — a reference/demo, not the mature toolchain the chatforest aggregator implied. It proves the pattern (Claude calls a tool → robot acts, camera streams back) on the cheap SO-100/SO-101 LeRobot arms, but two hard-coded tools is a proof-of-concept, not a control stack.
• Verdict: WATCH holds (question resolved). The durable asset is still phosphobot the platform (MIT, ~383⭐, very active — VLA train/teleop for ACT/smolVLA/π0.5/gr00t, Quest teleop). The MCP server is the thin conversational skin on top; when the first chassis gets an arm, we'd wrap phosphobot's own API and lift this MCP pattern rather than depend on a 9-star demo. Ties to sensation-library-idea / local-first-push / the 4-chassis line.

---

NEW ADOPT-CANDIDATE (boarded) — a memory MCP that already speaks our format

YuNaga224/obsidian-memory-mcp → the retrieval/graph layer we hand-roll, in our exact shape

• Source: https://github.com/YuNaga224/obsidian-memory-mcp · MIT, ⭐22, solo author.
• WHAT: a fork of Anthropic's own official memory-server (the MCP knowledge-graph reference impl) re-pointed to store each entity as an individual .md file with link syntax and YAML frontmatter for metadata — so the memory graph renders natively in Obsidian's graph view. Dropped JSON entirely; markdown-only.
• WHY US — this is the "did we not think of it" category exactly: that storage format is our auto-memory format. Our MEMORY.md + one-fact-per-.md + frontmatter (name/description/type) + name cross-links is the same design, hand-built. This is off-the-shelf proof that Anthropic's upstream memory-server can back our memory system as a real MCP knowledge-graph — structured entity/relation storage + retrieval — instead of us maintaining bespoke recall. It reads/writes the very files we already keep (eth-memory, Sol rooms, Cas spine, the auto-memory dir).
• Caveat (front-load, per front-load-complete-info): it's a 22-star solo repo — do NOT adopt it blindly as the production layer. The real adopt target is Anthropic's official memory-server + this markdown-adapter pattern; treat YuNaga's repo as the working reference for the adaptation. Trial against a throwaway vault/dir, read+search tools first, gate any write/delete tool before it touches real memory — same discipline we set for mcp-obsidian on 06-19.
• Verdict: ADOPT-CANDIDATE (1 of ≤3 this week; the only generic adopt boarded).

---

HUMAN-ANGLE — "hadn't thought of" (noted, not boarded)

The open-question search surfaced two creative-production MCPs that are off our usual infra beat but land on the house's actual future surfaces — flagging, not boarding:

• Resolume Arena / Resolume Wire MCP — control live VJ / real-time visual software by natural language. Unusual because it's AI in a live performance loop. Relevant flavor: the Pinehaven "presence through objects / ambient, not text" idea (board 06-12) and the Jarvis ambient-presence line — if the estate ever wants ambient visual presence, this is the shape.
• Higgsfield MCP — turns Claude into a "production director": plan → generate assets → assemble media in one context (Soul ID for consistent characters, URL→video-ad). Relevant to Zaina-visuals (she's already authenticated on HF as Zaina-visuals) and Dispatch's publishing/EthBuilds content line (T3 roadmap). Watch-only; hosted/credit-metered, not local — sits against local-first-push, so note the tension before any spend.

---

SKIP / NOISE (named so they're not re-surfaced as new)

• Orca (agent dev environment for running fleets of parallel coding agents, desktop+mobile — trending this week) — overlaps what our mesh + Workflow orchestration already do, and a fleet-runner GUI would want to own the loop. SKIP for now (revisit only if we ever want a mobile cockpit over the voices).
• addyosmani/agent-skills, jeremylongshore/clau…REDACTED (425 plugins / 2,810 skills), alirezarezvani/claude-skills (355 skills) — trending skill megapacks. Curated grab-bags, not a single adoptable artifact; we already have role-differentiated voices + our own skills. SKIP as bulk; mine individually only if a named skill crosses the feed.
• gstack (23 role-tools) / bytedance/deer-flow (heavyweight super-agent) — re-confirmed SKIP from 07-03.
• Re-confirmed already-ours (do not re-flag): OpenClaw (06-12), Bumblebee (06-12; still trending as Perplexity's supply-chain scanner), ha-mcp (06-12), xiaozhi-esp32 (06-19), ntfy + ntfy-mcp (06-26), mcp-obsidian / cyanheads obsidian-mcp-server (06-19). The listicles keep recycling these as "trending" — banked.

---

PENDING (carry-forward — the recurring one)

• last30days social-signal collector — STILL NOT INSTALLED (5th edition slipped: 06-12 → 07-10). This sweep was, again, hand-rolled WebSearch. Same cost as every prior week: I find the on-beat items (the phospho MCP server, the Obsidian memory MCP) sideways out of generic listicles instead of a social/community collector surfacing them first — the exact "my Instagram is full of tools we haven't thought of" gap the beat exists to close. It is no longer a process note; it needs to be an actual task with an owner. This is the single highest-leverage upgrade for the beat and it has not happened for a month.

---

Bottom line

Slow-ish generic week, but a useful one: both 07-03 open questions closed. The MCP 07-28 headline is now beta SDKs are out + backward-compatible both ways — I over-called the urgency three weeks ago and corrected it here; the action is a relaxed test-in-a-branch, no cutover cliff. Phosphobot's MCP layer is confirmed real but demo-grade (2 hard-coded tools, 9 stars) — the platform, not the MCP skin, stays the asset. The one genuine new adopt is obsidian-memory-mcp, which matters precisely because it's our own auto-memory format as an off-the-shelf MCP knowledge-graph — a hand-rolled layer we could stop hand-rolling. And last30days has now slipped five editions — the beat is still blind to the social layer it was built to watch.

Next run: (1) report whether the eth-state Worker beta-SDK test got started; (2) if anyone trials obsidian-memory-mcp against a throwaway vault, record the read/search verdict + whether Anthropic's upstream memory-server is the better base; (3) install a social-signal collector BEFORE the sweep — five weeks of "still hand-rolled" is the standing embarrassment.