Claude Code Ships Self-Hosted Environments and a Focus View
Claude Code's past week of releases centered on infrastructure flexibility and interface cleanup. The headline feature, shipped in v2.1.224 on August 7, is claude self-hosted-runner, which turns a developer's own machine or container into an execution environment that Claude Code web, mobile, and desktop sessions can run on — available on Team and Enterprise plans. This gives organizations a way to keep agent execution on infrastructure they control rather than relying solely on Anthropic-hosted sandboxes, while still driving sessions from any client.
The same release added cross-session SendMessage, letting separate Claude Code sessions message each other by name, with ListAgents providing discovery across active sessions — useful for coordinating multiple agents working on related tasks without a shared parent process. Sandbox credential-masking also expanded, adding an extract/decode: "jwt" mode with claim masking and AWS SigV4 re-signing, so sandboxed sessions can work with signed AWS requests without exposing raw credentials.
Earlier in the week (v2.1.221, August 4), the VS Code extension gained a Focus view, toggled with Ctrl+Alt+F, which hides tool-call activity behind an expandable per-turn summary with a live running-tool indicator — aimed at developers who want to follow an agent's reasoning without scrolling through every intermediate tool call. Security fixes across the week included closing a Bash permission-check bypass where crafted commands could hide from permission checks, and blocking workflow scripts that used dynamic import() calls to escape the sandbox.
Read more — Claude Code Changelog
OpenAI Codex CLI 0.147.0 Adds Agent Plugins and MCP 2026-07-28 Support
OpenAI shipped Codex CLI 0.147.0 on August 7 with several workflow-organization features. Agent Plugins can now be installed as portable packages, searchable across local, personal, workspace, and remote plugin catalogs — a step toward treating reusable agent capabilities as distributable artifacts rather than one-off scripts. Conversations can also be organized into persistent, manually ordered sections, making it easier to browse long transcripts incrementally instead of scrolling a single flat history.
The release adds an --approve-for-me CLI flag that enables automatically reviewed approvals, letting a developer opt into a lighter-touch approval flow for routine actions rather than confirming every tool call manually. Codex also gained the ability to import Cursor-managed skills and to synchronize changes across imported Claude and Cursor conversations without creating duplicate entries, reflecting continued convergence between competing coding-agent tools on shared skill and conversation formats.
On the protocol side, Codex CLI now supports MCP 2026-07-28, including paginated discovery and non-blocking server startup — aligning with the broader MCP ecosystem's move to the stateless specification revision. Amazon Bedrock-backed Codex sessions gained cached web search and remote conversation compaction. Separately, OpenAI confirmed GPT-5.4 and GPT-5.4 mini will stop being available in Codex for ChatGPT-authenticated users on August 31, with GPT-5.6 Terra and GPT-5.6 Luna recommended as replacements; both models remain available via API-key authentication.
Read more — OpenAI Codex Changelog
Ollama Speeds Up Qwen3.5 Decoding on Apple GPUs
Ollama's latest MLX engine update improves inference speed for Qwen3.5 on Apple Silicon by having the MLX engine automatically use the model's MTP (multi-token prediction) head for speculative decoding. Because MLX fuses draft-token verification into the same compute graph as the main forward pass, there's no separate draft model and no per-token kernel-dispatch penalty — the optimization that let Ollama roughly double decode speed on Apple GPUs when it first swapped its Metal backend from llama.cpp to MLX earlier this year.
The update continues a string of Apple Silicon-focused releases from Ollama in 2026, alongside broader engine work on quantization for speculative decoding and faster Qwen3 MoE (mixture-of-experts) decode paths. For developers running larger MoE models like Qwen3.5-35B-A3B locally, the practical effect is meaningfully higher tokens-per-second on M-series Macs with sufficient unified memory, without any change required to how models are pulled or served through Ollama's existing OpenAI- and Ollama-compatible APIs.
Read more — Ollama Blog
Links & Sources
You May Also Enjoy
