Claude Code v2.1.116 shipped April 20, 2026 with three real workflow improvements: /resume is 67% faster on sessions over 40MB, MCP tool results can now reach 500K characters instead of being silently truncated, and the thinking spinner shows inline progress messages. Each fix targets a specific friction point in daily production use.
I run Claude Code for hours every day -- 31 scheduled cron jobs and live interactive sessions. Version 2.1.116 is the kind of release that doesn't make headlines but makes the tool noticeably better. Here's what changed and why it matters for anyone using Claude Code seriously.
What Did v2.1.116 Actually Change?
Claude Code v2.1.116 ships five meaningful improvements: 67% faster /resume on large sessions, a 500K character cap for MCP tool results (up from a much lower default), inline thinking progress messages, deferred MCP resource loading on startup, and smoother terminal scrolling in VS Code, Cursor, and Windsurf. There's also a security patch preventing sandbox auto-allow from bypassing dangerous-path checks on rm and rmdir operations.
Released April 20, 2026, the update is available immediately via npm install -g @anthropic-ai/claude-code@latest or auto-update if you have it enabled. The full changelog is on the GitHub releases page.
The core improvements fall into two categories: performance (faster /resume, faster MCP startup) and visibility (thinking spinner granularity, immediate usage tracking). These aren't new features -- they're quality-of-life fixes for people who use Claude Code for hours at a stretch.
Why Was /resume Slow Before This Fix?
/resume is now up to 67% faster on sessions over 40MB. The root cause was inefficient handling of dead-fork entries during session file deserialization. The fix targets that path specifically, cutting the reload time on large sessions from 30+ seconds to under 10 seconds in common cases. Sessions under 40MB see minimal change -- the optimization targets the long-session pain point specifically.
Why do sessions get large so fast? Every file read, git diff, test run, and correction cycle adds tokens to the session file. A 4-6 hour development session with 50+ tool calls -- reading files, running tests, making edits, iterating on the output -- routinely hits 40-100MB. That's before you factor in extended thinking runs, which generate substantial reasoning traces. Multi-file refactors with test suites between each change push past 50MB in a single session easily.
I've hit this directly. Long refactor sessions where I'm fixing multiple interdependent files, running tests between each change, and correcting Claude's approach a few times -- those sessions accumulate fast. The 67% improvement is the difference between "this resumes fine" and "I'll just start over because the wait isn't worth it." That's a meaningful workflow change for anyone who leaves sessions open overnight or picks up interrupted work the next morning.
How Does the New 500K MCP Character Limit Work?
MCP tool results can now return up to 500,000 characters inline. Previously, results exceeding the default limit were silently truncated and persisted to disk with a file reference -- meaning Claude had to issue follow-up read commands to retrieve the full content, adding latency and occasionally losing context. Now MCP server authors can set anthropic/maxResultSizeChars up to the hard ceiling of 500K. If your tool returns a full result, it arrives intact.
The tools most affected by the old truncation behavior: database schema tools (full schemas for complex databases easily hit 50K+ characters), code review tools returning multi-file analysis with context, file tree operations with metadata, and deployment status reports covering multiple services. When results got truncated before, the broken data caused Claude to issue secondary retrieval commands -- adding round trips and sometimes producing incorrect analysis because it only had partial data to work with.
If you're building or maintaining MCP servers, this is the change to act on. Add anthropic/maxResultSizeChars: 500000 to your server config for any tool that generates comprehensive output. Don't leave it at the conservative default if your tool legitimately produces large results -- the new ceiling exists precisely so well-designed tools don't have to chunk their output artificially.
Want the templates from this tutorial?
I share every workflow, prompt, and template inside the free AI Creator Hub on Skool. 500+ builders sharing what actually works.
Join Free on Skool
What's New With the Thinking Spinner?
The thinking spinner now shows three distinct inline progress messages: "still thinking" for initial extended reasoning, "thinking more" when reasoning continues past the first checkpoint, and "almost done thinking" when Claude is wrapping up. Previously, a separate hint row showed a generic loading indicator with no state differentiation. The new messages are inline -- in the same response stream as the output -- and give you a real sense of where in the reasoning process Claude actually is.
This is a small change with outsized practical value. A generic spinner with no progress signal feels like a frozen process after 20+ seconds. These three states let you make a real decision: "almost done thinking" signals it's worth waiting; "still thinking" after 45 seconds tells you this is a deep task and you should go do something else. For /ultrareview runs on complex codebases, or extended thinking on architectural decisions, this visibility is genuinely useful rather than decorative.
What Are the Other Fixes Nobody's Talking About?
Beyond the headline improvements, v2.1.116 ships several changes worth knowing. MCP startup is faster because resources/templates/list now loads on first @-mention instead of at startup -- if you use multiple MCP servers, this reduces the initialization overhead on every session start. Terminal scrolling is smoother in VS Code, Cursor, and Windsurf via a new /terminal-setup scroll sensitivity configuration. /doctor can now be opened while Claude is actively responding -- no waiting for the current response to finish before running a health check. The /config search now matches against option values, so searching "vim" finds the Editor mode setting rather than requiring you to remember the setting name.
The security fix deserves a callout: sandbox auto-allow no longer bypasses dangerous-path checks for rm and rmdir. If you're running Claude Code in any automated context with sandbox settings, this closes a specific edge case where auto-allow could have been used to run deletions on protected paths. Standard interactive workflows are unaffected -- this targets a narrow but meaningful attack vector in automated environments.
Plugin auto-install also changed behavior: /reload-plugins and background auto-update now automatically install missing dependencies without manual intervention. For teams running Claude Code on multiple machines or in CI environments where plugins may not be pre-installed, this removes a common source of failure. Previously, you'd hit a plugin error and have to manually install the dependency -- now it handles it in the background.
How to Get the Most Out of v2.1.116 Right Now
Update first: run npm install -g @anthropic-ai/claude-code@latest and verify with claude --version. If you're on auto-update, you may already be on 2.1.116 -- but confirm, since the /resume improvement is specifically tied to this release. If you run Claude Code on multiple machines or in CI pipelines, update everywhere. The native binary from v2.1.113 plus the performance work in v2.1.116 is a meaningfully different runtime than what was available even two weeks ago.
For the MCP result size improvement to take effect, you need to update your MCP server config -- not just the Claude Code client. If you're using community or third-party MCP servers, check if they've added anthropic/maxResultSizeChars. If you're building your own servers, add it now. Set it to what your tool actually needs, not the ceiling. A tool that returns 50K results should be configured at 50K, not 500K -- match the setting to your tool's realistic output range.
For /resume: if you've been avoiding it on large sessions because of the slowness, try it now. The improvement is most pronounced on sessions in the 40-100MB range. If you tend to close sessions mid-task and restart fresh rather than resume, this fix changes that calculus. Combine /resume with /compact when sessions get long -- compact reduces the session size, resume loads it fast, and you preserve context without the multi-file context re-establishment overhead of starting fresh.
Where Does v2.1.116 Fit in the April 2026 Release Run?
April 2026 has been one of the most active months for Claude Code releases. Between v2.1.111 (April 16) and v2.1.116 (April 20), Anthropic shipped Claude Opus 4.7 integration with the xhigh effort level and task budgets, the /ultrareview command, a native binary replacing bundled JavaScript, sandbox network domain controls, and now the performance improvements in 2.1.116. Five meaningful releases in four days signals active investment in hardening Claude Code for production workflows, not just adding features.
The April timeline worth knowing: v2.1.111 added Opus 4.7 with task budgets, auto mode for Max subscribers, and /ultrareview for cloud-based multi-agent code review. v2.1.113 switched to a native Claude Code binary instead of bundled JavaScript and added sandbox.network.deniedDomains for network-level security controls. v2.1.116 is the polish release -- addressing performance debt from the rapid feature additions in the earlier builds.
For teams running Claude Code in CI/CD or scheduled automation, the April run matters as a whole. Native binary in v2.1.113 reduces startup overhead. Network sandbox in v2.1.113 enables stricter security policies. Performance improvements in v2.1.116 make long-session automation more viable. If you're building production agentic systems on Claude Code, you want all three -- not just the latest patch.
FAQ
How do I update to Claude Code v2.1.116?
Run npm install -g @anthropic-ai/claude-code@latest to update. Verify with claude --version -- look for 2.1.116 in the output. If you have auto-update enabled in Claude Code settings, it updates in the background between sessions. The v2.1.116 release shipped April 20, 2026 and is available immediately on npm.
What is the /resume command in Claude Code?
/resume reloads a previously closed or interrupted Claude Code session, restoring the full conversation and tool call history. Use it when you need to pick up mid-task work after closing your terminal. After the v2.1.116 fix, /resume is 67% faster on sessions over 40MB -- which covers most multi-hour development sessions with significant file editing and tool use history.
How do I set the MCP result size limit in my MCP server?
Set anthropic/maxResultSizeChars in your MCP server configuration to any value up to 500,000. If your tool returns large payloads -- database schemas, detailed reports, full file trees -- set this to avoid silent truncation. Without this setting, Claude Code uses a conservative default that truncates large results and persists them to disk instead of keeping them inline.
What is the xhigh effort level added in v2.1.111?
xhigh is a new extended thinking level between high and max, added in v2.1.111 alongside Claude Opus 4.7. It gives more reasoning depth than high without burning the full token budget that max requires. Use it for tasks where high-effort thinking isn't deep enough but max-effort is overkill or too expensive -- the sweet spot for complex architectural decisions or multi-step refactors.
Does the v2.1.116 sandbox security fix affect my existing workflows?
Only if you were relying on sandbox auto-allow to permit rm or rmdir on paths that would normally be flagged as dangerous. The fix ensures dangerous-path checks always run, even when auto-allow is enabled. Standard interactive and automated workflows are unaffected -- this closes a specific edge case where auto-allow could bypass deletion safety checks on protected directories.
Want the templates from this tutorial?
I share every workflow, prompt, and template inside the free AI Creator Hub on Skool. 500+ builders sharing what actually works.
Join Free on Skool
