Back to Blog
claude-code weekly-update auto-mode computer-use dispatch powershell security

What's Up Claude Code: Week of March 24th

Olivier Legris ·

TL;DR

The theme this week: Claude Code is becoming autonomous. Auto mode shipped as a research preview — a middle path between babysitting every command and --dangerously-skip-permissions. Computer Use landed on macOS, letting Claude control your screen, browser, and keyboard. Dispatch got even more useful when paired with both. Five releases in five days brought transcript search, PowerShell tool preview for Windows, conditional hooks, managed settings drop-in directories, and over 60 bug fixes. The community crossed 13,700 repos on GitHub, and the events map now spans 84 cities in 34 countries.

1. Auto Mode — The End of Permission Babysitting

Auto mode shipped on March 24 as a research preview for Team plan users (Enterprise and API coming soon). It’s the feature developers have been waiting for: Claude decides which actions are safe to run on its own, with a classifier screening every tool call before execution.

How it works: Before each action, a safety classifier checks for destructive behavior — mass file deletion, data exfiltration, malicious code execution, or anything the user didn’t request. Safe actions proceed automatically. Risky ones get blocked, and Claude redirects to an alternative approach. If Claude keeps hitting blocks, it escalates to a human prompt.

claude --enable-auto-mode
# Then cycle to auto mode with Shift+Tab

Why it matters: This is Anthropic’s answer to the “vibe coding babysitting” problem. You can kick off a large refactor, walk away, and come back to results — without the nuclear option of skipping all permissions. The classifier adds a small token/latency overhead, but removes the 50+ approval prompts that make long tasks painful.

The fine print: Auto mode works with Sonnet 4.6 and Opus 4.6 only. Anthropic recommends isolated environments. The classifier isn’t perfect — ambiguous user intent and unfamiliar environments can let risky actions through. But it’s a massive step up from the binary choice of “approve everything” vs. “approve nothing.”

TechCrunch covered it as part of a broader industry shift toward autonomous AI tools.

2. Computer Use — Claude Controls Your Screen

On March 23, Anthropic shipped Computer Use for Claude Cowork and Claude Code on macOS. Claude can now point, click, scroll, and navigate your screen when it doesn’t have a direct tool for the job.

The hierarchy: Claude reaches for the most precise tool first (connectors like Slack, Google Calendar), then falls back to direct screen control — opening files, using the browser, running dev tools. It always asks permission before accessing new applications.

Paired with Dispatch: This is where it gets interesting. Assign Claude a task from your phone via Dispatch, then Claude uses your computer while you’re away. Morning briefing while you’re on the train. IDE changes + tests + PR while you’re in a meeting. The phone becomes a remote control for your desktop AI.

Safety: Prompt injection detection runs automatically. Some apps are off-limits by default. You can stop Claude at any point. It’s a research preview — complex tasks sometimes need retries, and screen control is slower than direct integrations.

Available for Claude Pro and Max subscribers on macOS.

3. Transcript Search — Find Anything in Your Session

v2.1.83 added transcript search: press / in transcript mode (Ctrl+O) to search, then n/N to step through matches. For anyone running long sessions with hundreds of tool calls, this is a quality-of-life breakthrough. No more scrolling through 2,000 lines to find that one command output.

4. PowerShell Tool — Windows Gets First-Class Treatment

v2.1.84 added a PowerShell tool for Windows as an opt-in preview. This is the first native Windows shell integration — previously, Windows users relied on the Bash tool through Git Bash or WSL. Combined with improved dangerous command detection for Windows drive roots and better PATH inheritance for Git Bash in VS Code, Claude Code on Windows is becoming a real option.

Docs →

5. Conditional Hooks — Run Hooks Only When They Matter

v2.1.85 added an if field for hooks using permission rule syntax (e.g., Bash(git *)). Hooks only fire when the condition matches, reducing process spawning overhead. Combined with CwdChanged and FileChanged hook events from v2.1.83, the hook system is becoming a proper reactive event system — think direnv, but for your entire AI workflow.

6. Managed Settings Drop-in Directory

v2.1.83 added managed-settings.d/ alongside managed-settings.json. Separate teams can now deploy independent policy fragments that merge alphabetically. This is the enterprise feature that makes Claude Code manageable at scale — security, platform, and product teams each own their policy slice without merge conflicts.

Also new: CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips Anthropic and cloud provider credentials from subprocess environments (Bash tool, hooks, MCP servers). Defense in depth for managed deployments.

7. MCP & Remote Control Improvements

  • MCP OAuth follows RFC 9728 (v2.1.85) — Protected Resource Metadata discovery for finding the authorization server. The OAuth story is maturing.
  • MCP step-up authorization (v2.1.85) — Servers requesting elevated scopes via 403 insufficient_scope now correctly trigger re-authorization instead of failing.
  • MCP tool descriptions capped at 2KB (v2.1.84) — OpenAPI-generated servers were bloating context. Now they can’t.
  • Persistent ECONNRESET fix (v2.1.85) — Fresh TCP connections on retry, fixing a class of hanging MCP calls.
  • Remote Control now shows specific reasons when blocked, generates AI titles within seconds, and runs Esc-twice rewind picker in VS Code.

8. Performance & Memory

The week’s five releases collectively delivered significant savings:

ImprovementVersionSaving
--bare -p startupv2.1.83~14% faster to API
Startup with MCP serversv2.1.84~30ms + 600ms saved
Scroll stability (long sessions)v2.1.83Resets cut from 1/turn to 1/50 msgs
Large diff processingv2.1.835s timeout instead of hang
WASM → TypeScript layoutv2.1.85Better scroll + compaction perf
Non-streaming fallback capv2.1.8321K → 64K tokens, 120s → 300s
Read tool token usagev2.1.87Compact line format + dedup
Prompt cache (Bedrock/Vertex)v2.1.87Dynamic content removed from tool descs
Memory leak (long sessions)v2.1.87Render caches no longer retain full strings
Memory leak (remote sessions)v2.1.83Tool use IDs no longer accumulate

9. Security

  • sandbox.failIfUnavailable (v2.1.83) — Exit with error when sandbox can’t start, instead of running unsandboxed. Essential for CI where “silently unsandboxed” is a security incident.
  • --mcp-config policy enforcement (v2.1.83) — CLI flag was bypassing allowedMcpServers/deniedMcpServers. Fixed.
  • deniedMcpServers blocks claude.ai servers (v2.1.85) — Previously they slipped through.
  • Credential scrubbing (v2.1.83) — CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips provider credentials from subprocesses.
  • Plugin marketplace policy (v2.1.85) — Plugins blocked by org policy can’t be installed, enabled, or even seen in marketplace views.

10. Developer Experience

Small changes that compound:

  • Idle-return prompt (v2.1.84) — After 75+ minutes away, Claude nudges you to /clear instead of re-caching a stale session.
  • Token display (v2.1.84) — 1.5M tokens now shows as “1.5m” instead of “1512.6k.”
  • Interrupting before response (v2.1.83) — Your input is restored for editing. No more retyping.
  • /status during response (v2.1.83) — Works immediately instead of queuing until the turn finishes.
  • Plugin config (v2.1.83) — Plugins can prompt for configuration at enable time, with sensitive values stored in keychain.
  • Clipboard image paths (v2.1.83) — Claude can reference the on-disk path of pasted images for file operations.
  • [Image #N] chips (v2.1.83) — Pasted images insert positional chips so you can reference them in your prompt.
  • Background task notification (v2.1.84) — Tasks stuck on interactive prompts surface a notification after ~45 seconds.

11. Community Spotlight

13,703 public repositories now tagged claude-code on GitHub — up from 12,213 last week (+12% in one week).

Notable new projects:

  • harness (286⭐) — Meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use. Skills that create skills.

  • agentfiles (157⭐) — Discover, organize, and edit AI coding agent skills across Claude Code, Cursor, Codex, and Windsurf — from inside Obsidian. The skill IDE.

  • agentic-chatops (87⭐) — 3-tier ChatOps (n8n + GPT-4o + Claude Code) implementing all 21 patterns from “Agentic Design Patterns” — one operator managing 137 devices.

  • claude-meter (63⭐) — Track your Claude Code usage limits. Simple, useful, needed.

  • pentest-ai (52⭐) — Turn Claude Code into an offensive security research assistant. Specialized subagents for authorized pentesting, recon analysis, exploit research, and STIG audits.

  • glance (46⭐) — AI-powered browser automation MCP server for Claude Code. Navigate, click, screenshot, test — all from your terminal.

Trending this week:

  • everything-claude-code crossed 115K⭐ — the dominant agent harness
  • ui-ux-pro-max-skill hit 53K⭐ — design intelligence skill
  • oh-my-openagent (44K⭐) renamed from oh-my-opencode, broadening scope
  • cherry-studio (42K⭐) — AI productivity suite with 300+ assistants
  • cc-switch (35K⭐) — cross-platform launcher for Claude Code, Codex, OpenCode, OpenClaw & Gemini CLI

Pattern: The ecosystem is splitting into three tiers: harnesses that optimize how Claude Code runs (everything-claude-code, get-shit-done), orchestrators that coordinate multiple agents (ruflo, nanoclaw, dmux), and now meta-skills that generate other skills (harness, skill-creator). The abstraction ladder keeps climbing.

12. Community Events

155 events total (99 upcoming) across 84 cities and 34 countries.

This week (Mar 24–29):

Coming next week:

New cities this week: Atlanta 🇺🇸, Mérida 🇲🇽, Cancún 🇲🇽, Groningen 🇳🇱, Palo Alto 🇺🇸

Host your own on Luma →

What We’re Watching

  1. Auto mode adoption curve — Will teams trust the classifier enough to move off manual approvals? The classifier’s false-positive rate will determine adoption speed. Too many blocked benign actions = frustration; too few caught risky ones = incidents.

  2. Computer Use + Dispatch as the daily driver — The phone-as-remote-control pattern could fundamentally change how developers interact with their machines. Assign a task on the train, review results at the desk. Watch for “Dispatch-first” workflows emerging.

  3. PowerShell as a signal — Windows getting a native tool preview is Anthropic acknowledging the platform gap. If adoption is strong, expect Windows-first features. If not, it validates the macOS-dominant user base.

  4. Meta-skills and the abstraction ladder — harness, skill-creator, agentfiles — tools that generate tools. We’re entering the “frameworks for frameworks” phase. The question: does this create real leverage, or just more indirection?

  5. 13,700 repos and accelerating — +1,500 in one week. At this rate, claude-code becomes a top-50 GitHub topic before April ends. The ecosystem is past the tipping point — it’s self-sustaining.

  6. Five releases in five days — The team is shipping at a pace that suggests a major milestone on the horizon. Code with Claude conference (May 6 SF, May 19 London, June 10 Tokyo) is the likely reveal target.