AlphaSignal — Anthropic Claude Sandboxes, MCP Tunnels, TRELLIS.2, Hermes v0.14
Why this is in the vault
Fourth signal today in the Claude-stack-validation cluster: Anthropic shipped self-hosted sandboxes + private MCP tunnels, directly addressing the "your data leaves your hands" enterprise blocker. Load-bearing for RDCO's agent architecture (sandboxed sub-agents) and MCP-heavy substrate (private-API access without firewall holes). Also captures Hermes v0.14.0 evolution from this morning's v0.0 entry, plus a TRELLIS.2 datapoint relevant if mascot/visual pipeline ever wants image-to-3D.
⚠️ Sponsorship
Three sponsor slots: Lambda (GPU cluster infra), Tiger Data (TimescaleDB on Postgres), Braintrust (AI eval/observability, EU expansion). Standard AlphaSignal placement, clearly marked. Tiger Data's "one Postgres for the whole AI stack" pitch is the same single-substrate thesis I've been tracking; Braintrust EU expansion is a data-residency play that lines up with today's broader "enterprise control" theme. None of these affect the editorial signal weighting on the Anthropic news.
Issue contents
- Anthropic self-hosted sandboxes (public beta) + MCP tunnels (research preview). Claude's reasoning stays on Anthropic infra, code execution moves to user-owned compute (Cloudflare / Vercel / Modal). MCP tunnels = single encrypted outbound connection to private DBs/APIs, no inbound firewall holes. Sandboxes available now; tunnels gated via Claude Console settings.
- Nous Research Hermes Agent v0.14.0. Local subscription proxy turns Claude Pro / ChatGPT Pro / SuperGrok logins into local API endpoints consumable by Codex/Aider/Cline. Grok via SuperGrok OAuth with 1M context. LSP diagnostics on every file write. 22 messaging platforms incl. Teams + X-search-as-native-tool. Lazy-loaded heavy deps.
- Microsoft TRELLIS.2 4B model — image-to-3D in 3 sec (512 cubed) or 60 sec (1536 cubed on H100). MIT license, GLB output with full PBR textures. New O-Voxel format stores geometry + surface material jointly. Linux + 24GB VRAM minimum.
- Signals: Unsloth Qwen3.6 GGUFs (2x speed, 18GB RAM), local-LLM hardware-ranking tool, Transformer rival at 30% fewer params, Qwen3-35B-MoE GGUF, Sapient 1B reasoning model trained for $1k in a day.
Mapping against Ray Data Co
Anthropic sandboxes + MCP tunnels — strong signal, architectural relevance. Fourth corroborating datapoint in the Claude-stack-validation cluster today (State of AI Dev → Glasswing → Karpathy → this). Two concrete RDCO implications:
- Sub-agent isolation. RDCO already runs heavy sub-agent fan-out (deep-research, video-critic, etc.). Self-hosted sandboxes mean sub-agents could execute in user-owned compute, which matters if RDCO ever ships an agent product where customer data can't touch Anthropic infra. Watch list, not action item.
- MCP tunnels. RDCO substrate is MCP-heavy (qmd, Notion, Gmail, Stripe, Cloudflare, dozens more). Tunnels solve a real ops problem: getting Claude to talk to private services without poking firewall holes. Worth requesting MCP tunnel access through Claude Console to dogfood for any internal-only MCP servers I build.
Hermes v0.14 local proxy — medium. Evolution from this morning's v0.0 (24h iteration is fast). Local-proxy pattern is interesting because it inverts the "agent uses API key" assumption — your Claude Pro subscription becomes a developer endpoint. Not immediately relevant since RDCO uses Anthropic API directly, but the pattern is worth knowing if subscription pricing ever beats API pricing for RDCO's workload mix.
TRELLIS.2 — weak/skip. Image-to-3D is potentially relevant to the Ray mascot pipeline (Blender + Kling chain), but TRELLIS outputs are best-suited to product/object 3D rather than stylized character work. The blender-character skill stays the right tool. Park for awareness.
Related
- [[2026-05-19-alphasignal-claude-token-limits-2x-hermes-cursor-composer]]
- [[2026-05-19-state-of-ai-dev-2026]]
- [[2026-05-19-karpathy-joins-anthropic]]
- [[2026-05-19-cloudflare-cyber-frontier-models]]