Back to Blog
claude-code weekly-update mcp vscode partner-network context-window

What's Up Claude Code: Week of March 10th

Olivier Legris ·

TL;DR

The theme this week: Claude Code is becoming infrastructure for enterprises—not just a tool for developers. Five releases in five days drove home this shift: VSCode now shows Claude Code sessions in the activity bar, MCP servers can request structured input mid-task, Opus and Sonnet hit 1M context at standard pricing, and Anthropic announced a $100M Partner Network to help enterprises adopt Claude.

The message is unmistakable: Claude is scaling beyond early adopters. The infrastructure is hardening, the platforms are integrating, and enterprises are getting the support they need to move from POC to production.

Five releases (v2.1.72 → v2.1.76) shipped this week, headlined by VSCode revamp, MCP elicitation, and session naming. Plus a massive enterprise play: $100M invested in partners.

1. VSCode Integration Levels Up

v2.1.76 adds major VSCode improvements:

Sessions sidebar icon — Claude Code sessions now appear in the VSCode activity bar with a spark icon. Click it to see all your sessions, open full editors, and manage your work—no terminal needed. This is a UX shift: Claude Code stops being a sidebar utility and becomes a first-class workbench member.

Plan markdown editor — Plans now display as full markdown documents in VSCode with inline comments. No more squinting at JSON or plan trees—you get a readable, editable document right in your editor.

Native MCP management — Use /mcp in the chat to enable/disable servers, reconnect, and manage OAuth auth without touching the terminal. MCP integration moved from “power user” to “everyday.”

These three changes reframe Claude Code from an experimental CLI to a native IDE citizen. If you’ve been hesitant about CLI-only workflows, this week’s VSCode improvements are your on-ramp.

2. MCP Elicitation — Interactive Forms Mid-Task

v2.1.76 adds MCP elicitation: MCP servers can now request structured input from the user mid-task via an interactive dialog. Instead of Claude Code failing when it lacks a required parameter, the server can ask—and the user responds through a form or browser prompt.

New hooks:

  • Elicitation — intercept and override requests before they’re sent back
  • ElicitationResult — see what the user responded with

This is huge for stateful integrations. Imagine an MCP server for your payment processor: instead of failing when it doesn’t have an API key, it asks you for it on the spot. Or a database schema explorer asking you to authenticate before listing tables.

3. 1M-Context Opus & Sonnet — Standard Pricing, Prod-Ready

This week, 1M-context Claude Opus 4.6 and Sonnet 4.6 shipped at standard pricing—no upcharge for the extended context window.

  • Opus 4.6: 78.3% on MRCR v2 (highest among frontier models), 1M tokens
  • Sonnet 4.6: 1M tokens, priced in the mid-tier
  • Use cases unlocked: entire codebases in-memory, legal doc review, RAG over large datasets, research synthesis

This is the biggest context arbitrage in frontier models. Load your entire GitHub repo and get meaningful feedback. Load 500 pages of customer feedback and get a synthesis. This week marked the inflection point where “throw everything at the model” became economically viable for small teams.

4. Effort Level Control — /effort Command

v2.1.76 adds /effort to set model effort level. Opus 4.6 defaults to medium (the sweet spot: fast + thorough), but now you can dial it:

/effort high    # Use extended thinking
/effort medium  # Default
/effort low     # Fast inference

Return of ultrathink: The keyword is back—type it to bump effort for just the next turn without changing your setting.

This matters for cost-conscious workflows. You can set everything to low effort by default (cheaper, still smart), and /effort high when you hit a gnarly problem.

5. Sparse Checkouts for Monorepos

v2.1.76 adds worktree.sparsePaths — tell Claude Code which directories to check out via git sparse-checkout. For teams with massive monorepos:

{
  "worktree": {
    "sparsePaths": ["packages/api", "packages/web"]
  }
}

Only those paths get cloned. Saves time, disk, memory. Monorepo complexity stops being an excuse to avoid Claude Code.

6. Session Naming & PostCompact Hook

-n / --name flag at startup: claude --name "MyProject" gives you a display name without having to rename the session mid-chat.

PostCompact hook — fires after compaction completes. Use it to log, clean up, or notify. Useful for agents that need to know when the brain was compressed.

7. Anthropic Invests $100M in Claude Partner Network

The week’s biggest news: Anthropic announced $100M commitment to the Claude Partner Network—a program for partners helping enterprises adopt Claude.

What it includes:

  • Training courses + sales enablement
  • Joint market development + co-marketing
  • Technical support (dedicated Applied AI engineers, architects)
  • Certifications (first: Claude Certified Architect, Foundations)
  • Code Modernization starter kit
  • Services Partner Directory (where enterprises find implementation firms)

Why it matters: Enterprise adoption isn’t bottom-up. It’s top-down and requires trusted guides: McKinsey, Accenture, Deloitte, Infosys, Cognizant. These firms handle compliance, change management, integration. Anthropic just signaled it’s serious about supporting them.

The quote that stuck: “Anthropic is the most committed AI company in the world to the partner ecosystem—and we’re putting $100 million behind that this year to prove it.” — Steve Corfield, Head of Global Business Development, Anthropic

This is OpenAI playing offense in enterprise; Anthropic playing defense and building a fortress. The partnership push suggests Anthropic sees enterprise support as a moat against commodity LLM competition.

8. Community Projects This Week

Top releases by stars (Mar 10-15):

Garry Tan’s gstack (11.6K⭐) — Y Combinator president’s personal Claude Code setup: six opinionated tools + configuration. Clear signal that Claude Code + curated tools = productivity engine.

Moltbrowser MCP — Browser automation as an MCP. Community pick of the week: “Coolest project I’ve seen this week.”

Agency Agents — Multi-agent orchestration framework. “Pretty cool” first impressions.

Audio Analyzer (Rust) — Low-level audio processing skill for Claude Code.

MCP2CLI — Bridge MCP servers to CLI tools. Useful for non-integrated environments.

Paperclip — “Control center with employees” — orchestrate multiple Claude Code sessions like a team.

Tadabur — AI research assistant skill.

Other highlights:

Pattern: Skills are becoming the primary way people extend Claude Code. Not hooks, not plugins—skills. Encapsulated, shareable, composable.

Media:

9. Bug Fixes & Stability

Major themes this week: VSCode polish, network edge cases, permission clarity.

VSCode fixes:

  • Spark icon for sessions management
  • Plan markdown editor
  • Native MCP dialog UI (no terminal needed)
  • Fixed plugin installations lost in multi-instance runs

Network & Auth:

  • claude.ai OAuth reconnection failures fixed
  • Improved error messaging for microphone silence (distinguishing from “no speech detected”)
  • Reduced Remote Control /poll rate from 1–2s to once per 10 minutes (300× server load cut)

Permissions:

  • /rename now works while Claude is processing (was silently queued)
  • Reduced prompt re-renders by 74% during turns (performance win)
  • Startup memory down ~426KB (CA cert handling)

10. What We’re Watching

  1. MCP elicitation adoption — Will stateful server integrations become the new standard? (Probably yes within 6 months.)

  2. MCP renaissance — Moltbrowser, Paperclip, MCP2CLI signal MCPs are becoming the glue between Claude and existing tools. The “MCP is dead” narrative was premature.

  3. 1M-context usage patterns — Will devs actually load entire codebases, or stick to chunking? (Initial signals: yes, for monorepos + RAG.)

  4. Partner Network velocity — How many enterprises actually move through the partner program to production? (Anthropic’s bet: this is the choke point for 10%+ adoption.)

  5. VSCode as primary UX — Does this finally make Claude Code native to Microsofties? (Yes. Terminal-phobic teams can now onboard.)

  6. Agency patterns — Are we moving from “Claude Code runs alone” to “Claude Code runs other agents”? (Yes, see Paperclip, Agency Agents.)

Community Events

34 events upcoming across 6 continents. Key dates:

This week (Mar 15-21):

Coming:

Host your own on Luma →