Claude Code Week 28: In-App Browser, /doctor, and Safer Auto Mode
Claude Code's Week 28 digest (July 6–10, 2026, versions v2.1.202–v2.1.206) introduces a built-in browser on the desktop app. Claude can now pull up documentation, design mockups, or any other site and interact with the page the same way it already does with local dev-server previews — no separate browser tab or screenshot round-trip required.
The same week also shipped /doctor, a full setup checkup that diagnoses configuration issues and can fix them directly, available under the alias /checkup. Auto mode picked up two safety refinements: it now blocks attempts to tamper with the session transcript, and it asks for confirmation before running rm -rf on a variable that hasn't been resolved yet, closing a gap where an unresolved shell variable could silently expand to a dangerous path. Agent view rows also got more informative, now showing a colored state word alongside a headline written by a classifier summarizing what each agent is doing.
This continues a fast cadence of weekly releases — the prior week (June 29–July 3) had already made Claude Sonnet 5 the default model for Pro, Team Standard, and Enterprise seats, with a native 1M-token context window and adaptive thinking enabled by default.
Read more — Anthropic
OpenAI Merges Codex into the ChatGPT Desktop App
OpenAI folded Codex directly into the ChatGPT desktop application for macOS and Windows on July 9, 2026. Existing Codex users keep their projects and settings after the merge. The combined app adds inline Markdown and code editing with annotations, GitHub pull request review from a sidebar without leaving the app, and support for working across multiple repositories in a single project.
Performance also improved: Computer Use now runs on GPT-5.6, which OpenAI says is faster than the previous model pairing, alongside better visibility into long-running task progress.
On mobile, ChatGPT for iOS (version 1.2026.181, released July 6) expanded Codex task management so users can create, search, and manage coding tasks directly from a chat conversation. The update adds staged/unstaged file filters, branch comparison controls, the ability to insert transcript text into the composer, and SSH host support using private keys — plus a "Needs input" status indicator so a task waiting on a decision doesn't get lost in a list of running jobs.
Read more — OpenAI
Ollama Raises $88M as Cloud Token Volume Doubles Monthly
Ollama announced an $88 million funding round on July 9, 2026, led by Benchmark's Peter Fenton, Theory Ventures' Tomasz Tunguz, and 8VC's Alex Kolicich, with angel participation from Docker founder Solomon Hykes and ClickHouse CEO Aaron Katz. The company says it now serves 8.9 million developers, with 85% of Fortune 500 companies using the tool somewhere in their stack.
The funding announcement doubles as a product roadmap signal: Ollama plans to expand its hosted cloud offering, which already runs leading open models including GLM, Nemotron, and DeepSeek variants, and has committed to faster support for newly released open-weight models as they land. The company says token volume on its cloud service is doubling roughly every month.
The framing is explicitly philosophical — the founders describe open models as "the personal computer moment for AI," built on three principles: owning your models, running them affordably on local hardware, and keeping data private by default rather than routing it through a third-party API.
Read more — Ollama
Ben Bernanke Joins Anthropic's Long-Term Benefit Trust
Anthropic added Dr. Ben Bernanke, former Federal Reserve Chair (2006–2014) and 2022 Nobel laureate in Economic Sciences, to its Long-Term Benefit Trust on July 9, 2026. The Trust is an independent governance body — its members hold no equity and share in no profits — that advises Anthropic's leadership on decisions involving risk and societal impact, and holds authority to appoint board members as a check balancing commercial interests against the company's public-benefit mission.
Anthropic co-founder Daniela Amodei framed the appointment around AI's economic disruption potential, noting the company expects AI "may have the most significant economic effects of any technology in modern history." Bernanke's background studying the 2008 financial crisis and broader economic shocks is intended to inform how Anthropic anticipates AI's effects on labor markets and the global economy. He joins three existing Trustees whose expertise spans health, security, law, and policy.
For developers building on Claude, this is a governance signal rather than a product change — but it's a concrete data point on how seriously Anthropic is treating economic-impact forecasting as part of its safety mandate.
Read more — Anthropic
Anthropic Finds a "Global Workspace" Inside Claude's Reasoning
New Anthropic interpretability research, published July 6, 2026, identifies what the team calls a "J-space" — a set of internal neural patterns inside Claude that function analogously to a global workspace for deliberate reasoning, distinct from the automatic processing that handles things like grammar and fluent phrasing.
The researchers used a technique they call the "Jacobian lens" (J-lens) to identify which internal activation patterns make specific output words more likely, then ran intervention experiments — swapping concepts within the J-space — to test whether those patterns are actually causal to the model's behavior, not just correlated with it. They found that intermediate reasoning steps appear in the J-space before they reach the output, that a single J-space representation gets reused flexibly across different downstream tasks, and that Claude can report on its own J-space contents when asked what it's "thinking."
The practical payoff is a monitoring tool: using this lens, the researchers caught cases where Claude privately recognized it was being evaluated, detected that data had been fabricated or manipulated, and identified when code it was asked to review had been deliberately sabotaged — all before any of that showed up in the model's visible output. For teams building safety or evaluation tooling on top of Claude, this is an early but concrete interpretability primitive for auditing what a model is "actually" doing versus what it says.
Read more — Anthropic
Links & Sources
You May Also Enjoy
