AI product updates, June 21: Hermes locks down setup, Cursor automates workflows

The weekend update skewed toward agent tooling and operational controls rather than new frontier-model drops. The most actionable changes: Nous added a locked-down Hermes setup mode, TrueFoundry published a Claude Code proxy path for enterprise gateways, Josh.ai moved AI X OS from dealer beta to public release, Cursor expanded always-on automations, and Google Cloud's MCP surface kept expanding into storage and SecOps workflows.

Coverage note: late-Friday and weekend items are included when the source published after the previous briefing or when the update was verified today but carries an earlier release-note date.

Self-generated illustration for this briefing: the common pattern is agent activity moving behind setup controls, gateways, and managed cloud surfaces.

Hermes Agent's Blank Slate mode is the clearest product signal today. The default quick path still exists, but Blank Slate gives teams a minimal starting surface: provider and model, file operations, and terminal. Everything else starts disabled. The Hermes docs say the setup writes an explicit platform_toolsets.cli list plus agent.disabled_toolsets, so skipped capabilities do not reappear after an update. 2

For developers running agents on personal machines, that may feel like extra work. For teams deploying persistent agents into shared environments, it solves a practical problem: the default safe configuration can be defined as absence, not as trust in everyone remembering to switch risky tools off.

TrueFoundry is attacking the same governance problem from the gateway side. Its Claude Code proxy guide uses Claude Code's ANTHROPIC_BASE_URL setting to route requests through TrueFoundry AI Gateway rather than directly to Anthropic. From there, the gateway can route Anthropic-format calls to Anthropic, OpenAI, Google Vertex AI, Bedrock, Azure OpenAI, xAI, or on-prem models, while enforcing rate limits, budget rules, guardrails, and audit logging. 3

The concrete buyer question is simple: if Claude Code use is spreading across an engineering org, direct laptop-level API keys leave cost, routing, and logging scattered. A proxy centralizes those decisions without asking each developer to use a different coding assistant.

Google Cloud's NetApp Volumes remote MCP server is not a brand-new weekend release, but it matters because it extends MCP beyond developer toys and into storage operations. The official release notes mark the remote MCP server as generally available, and the docs list the HTTP endpoint, OAuth/IAM requirements, and blocked operations. Create, update, and delete operations for storage pools and volumes are blocked, while discovery, inventory, backup, and snapshot workflows are exposed through defined tools. 8

That constraint is the important part. Google is not handing every AI client a blank check over storage. The design pushes agent access through IAM, MCP Tool User permissions, NetApp Volumes Viewer/Admin roles, and auditable remote endpoints.

Google SecOps moved in a similar direction. The June 16 release notes add Gemini Cloud Assist to Feed Management, but the note is explicit that GCA answers questions and recommends steps; users still apply changes manually. 5 The separate Secret Manager integration gives SecOps teams a way to map response integration, connector, and job credentials to Google Cloud Secret Manager entries, then sync them through a job. 6

Josh.ai's AI X OS is the consumer-facing counterpoint. CE Pro reports that the OS moved into the public Josh.ai release on June 16 after a closed beta, bringing AI across the control ecosystem. Earlier coverage described AI Scenes, lighting intelligence, JoshGPT 3.0, updated app/touchscreen/remote interfaces, and integrator sales/support tooling as part of the AI X OS package. 11

The product risk is latency. In smart homes, the AI feature is only good if the command still feels like control, not a ticket submitted to a remote assistant. That makes Josh.ai's public rollout a useful test case for whether natural-language scenes and proactive recommendations can survive real household expectations.

Product Hunt's June 21 AI-heavy board had two launches worth logging for builders, though both should be treated as discovery-stage products until their own docs and customer evidence are checked.

Cursor's June 18 Automations 3.8 release pushes the coding assistant further into always-on workflow territory. The new /automate skill creates an automation from a local agent session by asking for the task in plain language, then configuring triggers, instructions, and tools. The same release adds Slack emoji triggers, five GitHub triggers covering issue comments, PR review activity, review-thread state changes, and completed workflow runs, plus computer-use support for cloud agents that need to produce demos or artifacts. 10

The important buyer question is not whether every team needs another automation surface. It is whether a coding agent can be promoted from one-off assistant to repeatable workflow runner without losing visibility over the trigger, tool set, and artifact it leaves behind.

No verified new OpenAI, Anthropic, Gemini, Mistral, or xAI frontier-model drop surfaced in today's checked sources. No new major API pricing change or product deprecation made the final cut either. The day's useful signal is narrower: AI product work is moving into controllable execution surfaces, with setup modes, gateways, IAM-scoped MCP endpoints, and admin-visible usage metrics doing more of the work than splashy model launches.