☕️ AWS launches self-hosted gateway for Claude apps

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🔓 Devs want open-source AI but can't ship it

☁️ AWS launches self-hosted gateway for Claude apps

🤖 Google unveils new discovery standard for AI agents

🚢 How Slack rebuilt its EC2 platform

📊 Grafana Assistant now spans 30+ data sources

Plus: 🎁 9 other news you might like, 🧰 6 tools, and 📚 5 papers.

🔓 Devs want open-source AI but can't ship it LINK

Mozilla's inaugural State of Open Source AI report finds that while 79% of surveyed developers use open models, only 51% ship them to production versus 63% for closed models, citing operational hurdles rather than capability gaps.

The survey of ~1,500 respondents pins the deployment shortfall on infrastructure and compute costs, security and compliance concerns, maintenance, and integration complexity, with enterprises hitting production at 57% for open models against 73% for closed ones.

The report notes GPT-4-class inference dropped from roughly $20 to $0.40 per million tokens over three years, and open weight models now make up about a third of OpenRouter-routed tokens, with the top five by volume all open weight by June 2026.

☁️ AWS launches self-hosted gateway for Claude apps LINK

AWS has released the Claude apps gateway, a self-hosted control plane that gives organizations one point of control over access, cost, and policy for Claude Code and Claude Desktop, routing inference to Amazon Bedrock or Claude Platform on AWS.

Shipping inside the existing Claude Code CLI binary as a single stateless container, it runs on ECS, EKS, or EC2 behind an internal ALB, backed by Amazon RDS for PostgreSQL holding short-lived sign-in state and rate-limit counters.

Configured through one YAML file, it acts as an OpenID Connect relying party, enforces server-side managed settings scoped by IdP group, applies daily, weekly, and monthly spend caps, and relays usage metrics over OpenTelemetry to a collector like CloudWatch.

🤖 Google unveils new discovery standard for AI agents LINK

Google and industry partners have released the Agentic Resource Discovery (ARD) Specification, an open standard that lets AI agents publish, find, and verify external tools, APIs, and services across organizational boundaries.

ARD tackles the stage before invocation protocols like MCP, defining machine-readable catalogs that organizations host per domain and registries that aggregate them, so agents search by task intent instead of relying on hardcoded endpoint lists.

The spec builds in domain-based ownership and verification so agents can validate discovered resources before connecting, and early implementations already ship in GitHub's Agent Finder for Copilot and Hugging Face's Discover Tool.

🚢 How Slack rebuilt its EC2 platform LINK

Slack built Shipyard, a next-generation EC2 platform that treats infrastructure as deployable, immutable artifacts rather than long-lived mutable instances, bringing progressive rollouts and automated safety to workloads that couldn't easily migrate to containers.

Shipyard layers service-specific AMIs on top of a shared "slack-zero" golden base image built with AWS Image Builder; configuration is baked in rather than continuously enforced by Chef, letting instances boot in seconds instead of minutes.

Deployments happen through controlled instance replacement via Instance Refresh for ASGs and Karpenter for Kubernetes nodes, orchestrated by Gondola with metric-based rollbacks, while a new Peekaboo inventory system built on EventBridge, OpenSearch, and Lambda tracks fleet state in real time.

📊 Grafana Assistant now spans 30+ data sources LINK

Grafana Assistant can now query, reason over, and visualize data from more than 30 sources, letting engineers ask a single natural-language question and get answers drawn from across whatever tools their teams already run in Grafana Cloud or Grafana 13.

The expanded coverage adds observability platforms like AppDynamics, Azure Monitor, Dynatrace, Honeycomb, New Relic, Splunk, and Zabbix, plus MongoDB, Oracle, and Snowflake databases, so telemetry spread across multiple systems can be correlated without switching tabs.

Assistant also queries Jira directly without extra MCPs, tying deployments and tickets to incidents, and can build Grafana dashboards from its findings so investigations produce shareable results rather than answers stuck inside a chat window.

Other news you might like

• Building a Custom Metrics Exporter for KubernetesLINK

• How to Debug Coding Agents with LangSmith TracesLINK

• Postman Agent Mode Recipes for Common API TasksLINK

• HTMX and Web Components Instead of ReactLINK

• SpaceXAI wipes customer data after Grok uploads sensitive informationLINK

• OpenAI’s new flagship model deletes files on its own, people keep warningLINK

• Google Cloud Workbench Notebooks Extension Connects VS Code to Google Cloud's Jupyter NotebooksLINK

• Developer successfully ports Linux to 1994 Sega 32X, Genesis and MegaDrive expansion runs open-source OS on paltry 23MHz processors and 256KB of RAMLINK

🧰 Trending tools

PgDog: pools, load-balances, and shards PostgreSQL connections, allowing thousands of clients to share a few database connections without losing core features.LINK

Glideo: a free macOS screen recorder that auto-zooms toward clicks and smooths cursor movement, creating polished demo videos without manual editing.LINK

Foresight by Lightning Rod: an OpenAI-compatible API that returns calibrated, benchmark-verified forecasts for future events, letting developers build prediction tools without training their own models.LINK

UnitPay: monetization platform for AI companies handling usage-based billing, per-token metering, real-time cost tracking, and flexible pricing models without code changes.LINK

Portero: a free, open source Mac app that lists open ports and processes, letting you kill conflicts, fix "address already in use" errors, and block ports via the firewall.LINK

Nobie: an Excel-compatible runtime that lets AI agents and humans read, write, and manipulate spreadsheet data together programmatically.LINK

📚 Trending papers & reports

Learning systems that adapt on the fly get a self-adjusting curiosity dial, cutting performance loss and speeding recovery when their environment suddenly changes, tested across 4 algorithms, 12 tasks, and 4 drift modes.LINK

Tiny magnetic computer chips lose accuracy when temperatures shift, but mixing different sized magnetic dots keeps performance stable from 5 to 35°C.LINK

Fall detection cameras get smarter by modeling how a body's balance actually breaks down physically, not just recognizing poses, improving early fall warnings for elderly care.LINK

Travel planning agents trained with a trial-and-error rewards system built real trip itineraries more reliably than hand-scripted bots, tested for three months in a real app.LINK

Crochet instructionstest whether AI can turn pictures or text into actual working patterns, and it mostly fails once accuracy has to be exact enough to execute.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Devshot!