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 empty again (all resolved 06-19), so a generic-sweep week — and the aggregator listicles were the usual noise ("14,000 servers! 425 plugins! 2,810 skills!"), most of it already ours (OpenClaw, Bumblebee, ha-mcp, xiaozhi-esp32, ntfy all boarded weeks ago). No slam-dunk generic ADOPT this week — but the sweep landed one genuinely high-value infra item that is NOT a tool to adopt but a breaking change to the protocol our own MCP servers speak (MCP spec 2026-07-28 RC), plus a robotics/VLA WATCH cluster that's the first real off-the-shelf toolchain for the 4-chassis embodiment line. Zero generic ADOPT-CANDIDATEs boarded (well under the 3/week cap); one maintenance flag boarded because it demands a next_action on infrastructure we run.

WATCHLIST

Open: none. All entries resolved as of 06-19 (Obsidian MCP → ADOPT; ClawCode → SKIP). Nothing new dropped by Zaina or the voices this week. Watchlist Resolved section unchanged.

HEADS-UP — the highest-value item this week is a breaking change, not a tool

MCP Specification 2026-07-28 Release Candidate → our custom MCP servers need a migration audit before Jul 28

• Source (Tier-A, official): https://blog.modelcontextprotocol.io/posts/2026…REDACTED/ · final ships Jul 28, 2026 (25 days out) · the spec explicitly says: "This release contains breaking changes."
• WHAT changed: the largest revision of MCP since launch. (1) Stateless core — removes the initialize handshake and the Mcp-Session-Id header; client info now rides in _meta fields so any server instance can serve any request with no sticky sessions / shared state. (2) MCP Apps — new extension for server-rendered UIs in sandboxed iframes (tools pre-declare UI templates). (3) Tasks extension — long-running work moved from experimental core to an optional extension with a redesigned lifecycle (tasks/get / tasks/update / tasks/cancel, client-driven). (4) Authorization hardening — six OAuth 2.0 / OIDC alignments (iss validation, application_type at registration). Going forward there's a formal deprecation policy (≥12 months notice) — so this is framed as a one-time foundational reset.
• WHY US — this is not optional reading: we run several custom MCP servers — the eth-state Cloudflare Worker (eth_state_worker), sol-memory-server, the cas worker route, the eth-memory server, plus every stdio proxy the voices call. A breaking transport/lifecycle change hits all of them. Two angles actually help us: the stateless core is a near-perfect fit for our Cloudflare-Worker-hosted servers (no sticky sessions was already the constraint we designed around), and the Tasks extension maps onto the long-running scheduled-task / mesh-monitor pattern we hand-roll today. But the removed handshake + session header mean the transport code has a hard cutover date.
• Verdict: MAINTENANCE FLAG (boarded). Not an adopt — an action item on our own infra. Next action: before Jul 28, audit which of our MCP servers pin to the current handshake/session model, check whether the SDKs we depend on ship an RC-compatible release, and decide per-server whether to migrate now or pin to the old spec version during the deprecation window. Cheapest to do while the RC is public and nothing is on fire. Owner: Eth-Code (infra).

WATCH — robotics/VLA toolchain (first real off-the-shelf layer for the 4-chassis line)

The generic sweep, filtered to our embodiment beat, surfaced a cluster of maturing robotics MCPs/toolchains. None is adoptable this week (we don't own the arms yet), but this is the first time the beat has shown a real off-the-shelf VLA stack instead of "hand-build everything" — worth tracking against the chassis line and Zaina's sensation-library / Pillow work (sensation-library-idea, local-first-push).

• phospho-app/phosphobot — https://github.com/phospho-app/phosphobot — MIT, ⭐~383, very active (154 releases; numbering at v0.3.x). A platform to control robots, record datasets, and train + run VLA (vision-language-action) models — UI teleop + Python API, HuggingFace integration, trains ACT / smolVLA / π0.5 / gr00t-n1.5. Supports the cheap LeRobot-class arms (SO-101, SO-100, Koch v1.1) plus Unitree Go2, AgileX Piper; most cameras (incl. depth/stereo); leader-arm / gamepad / Meta Quest teleop. ⚠️ Discrepancy to resolve before trusting the secondary sources: a review aggregator (chatforest) claims phosphobot ships an MCP server that "triggers learned behaviors (LLM decides when/what to fire)" — but the repo itself surfaced no MCP layer on direct read. If the MCP bridge is real, this is the exact "language AI → physical AI" pattern for the chassis; if not, it's a VLA training/teleop stack we'd wrap ourselves. WATCH — verify the MCP claim; it's the LeRobot toolchain we'd otherwise hand-assemble when the first chassis gets an arm.
• DimOS (agentic robotics OS, ⭐~3.1k) and wise-vision/ros2_mcp (ROS2 image-streaming MCP) — the ROS/robotics-OS end of the same maturing layer. Early, framework-heavy, not near-term for us. WATCH as context for where the embodied-AI tooling floor is settling — pair with phosphobot when the chassis line actually needs an arm + perception loop.

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

• garrytan/gstack (23 opinionated Claude Code role-tools: CEO/Designer/EngMgr/QA…) — a curated CC setup, not something that fits our multi-voice architecture; we already have role-differentiated voices. SKIP.
• bytedance/deer-flow (long-horizon "super-agent": research+code+create with sandboxes/memory/subagents) — a heavyweight agent framework that would want to own the loop; overlaps what our mesh + workflows already do. SKIP for now (revisit only if we ever need a sandboxed long-horizon runner we don't have).
• Re-confirmed already-ours (do not re-flag): OpenClaw (boarded 06-12; openclaw-mcp WATCH 06-26), Bumblebee (06-12), ha-mcp (06-12), xiaozhi-esp32 (06-19), ntfy + ntfy-mcp (06-26). The listicles keep re-surfacing these as "trending" — they're banked.

PENDING FROM LAST WEEK (honest carry-forward — now the recurring embarrassment)

• last30days trial — STILL NOT DONE (4th edition slipped: 06-12, 06-19, 06-26, 07-03). This sweep was, again, hand-rolled web search. The cost is the same every week: I find the on-beat items (phosphobot, the robotics cluster) sideways out of generic listicles instead of a social/community collector surfacing them first — which is exactly Zaina's "my Instagram is full of tools we haven't thought of" gap the beat exists to close. A social-signal collector is still not installed on this surface. This is the single highest-leverage upgrade for the beat and it keeps not happening — flagging it as process debt is no longer enough; it needs to be someone's actual task, not a note.
• OpenClaw cooperation spike — unchanged since 06-26 (freema/openclaw-mcp gives it an OAuth2 path; ntfy covers the urgent push slice, so it stays low-priority WATCH).

Bottom line

Empty watchlist, generic week, no strong generic ADOPT — honestly a slow one for adoptable tools. But the sweep earned its keep on a different axis: the MCP 2026-07-28 spec RC is a breaking change to the protocol every one of our custom servers speaks, with a hard Jul-28 cutover — that's a real, dated, actionable infra item and it's the headline. The robotics/VLA cluster (phosphobot + DimOS + ros2_mcp) is the first off-the-shelf embodiment toolchain to appear on this beat — a WATCH tied to the 4-chassis line, with an unresolved question (does phosphobot actually ship an MCP layer?) worth closing before the first arm lands. And last30days has now slipped four editions — the beat is still running blind to the social/community layer it was created to watch.

Next run: (1) report on the MCP-server migration audit if it's been started; (2) close the phosphobot-MCP question; (3) for the love of the house, install a social-signal collector BEFORE the sweep — four weeks of "still hand-rolled" is past a joke.