Fixing slow Zed agent startup times

17 Dec 25

aidev

I’ve been using Claude Code via ACP (Agent Control Protocol) in Zed, and while the integration is brilliant, starting a new Claude session took noticeably longer in Zed than running claude directly in the terminal; ~30s vs about 5s. And it was particularly frustrating because there’s no /clear slash command via ACP—the only way to clear the context, which I d all the time is to re-start the agent, which (because of the delay) breaks my flow.

Digging into the diagnosis

Looking into it, issue #13329 documents how MCP server initialisation blocks session creation. One user had measured their setup taking 32+ seconds just waiting for MCP servers to initialise. Seems like just the issue I was having.

So I checked my local MCP configuration. I have a few servers “globally” configured for all projects on this machine:

I ran some timing tests and found the smoking gun: @playwright/mcp was being invoked via npx @playwright/mcp@latest, so it checks the npm registry every time and potentially downloads packages. Meanwhile, the other MCP servers (backlog and ht-mcp) were fast, because they were already installed globally and being invoked directly as binaries. Playwright was the outlier, and it was dragging down every single session startup.

The fix

I implemented a three-part fix:

1. Install Playwright MCP globally

Instead of relying on npx to fetch it every time, install it globally:

npm install -g @playwright/mcp

2. Add mise shims to PATH for non-interactive contexts

Here’s where things got interesting. My mise setup was working fine in the terminal (where I have mise activate zsh in my .zshrc), but IDEs like Zed don’t run interactive shells—they use non-interactive contexts that skip .zshrc entirely1.

The solution is mise’s recommended hybrid approach: add the mise shims directory to PATH in .zshenv (which is sourced for all contexts), while keeping mise activate in .zshrc for interactive shells:

# .zshenv (sourced for all shell contexts)
export PATH="$HOME/.local/share/mise/shims:$PATH"

# .zshrc (sourced for interactive shells only)
eval "$(mise activate zsh)"

This ensures that globally-installed npm packages are available to MCP servers spawned by Zed, without sacrificing the nice interactive shell features that mise activate provides.

3. Update MCP config

Finally, update the MCP configuration to use the binary directly instead of going through npx:

{
  "mcpServers": {
    "playwright": {
      "command": "mcp-server-playwright"
    }
  }
}

The outcome

The result? It’s still not instant, but it is faster. The pause that was breaking my flow is (closer to) gone.

The broader lesson here is that npx is convenient for one-off commands, but it’s a terrible choice for anything that runs frequently. Global installation plus direct binary invocation is the way to go for tools you use regularly2. And if you’re using mise (or any other tool version manager like asdf or rtx), remember that non-interactive contexts need explicit PATH configuration—IDEs and MCP servers don’t get the benefits of your shell’s interactive setup automatically.

Footnotes

  1. This is standard Unix behaviour. Interactive shells source .zshrc (or .bashrc for bash), but non-interactive shells only source .zshenv (or .bash_profile). IDEs, MCP servers, and other tools that spawn processes typically use non-interactive contexts. ↩

  2. The trade-off is that you now need to update these packages manually (npm update -g @playwright/mcp) rather than getting the latest version automatically. For rapidly-evolving tools that’s a real consideration, but for stable packages the performance gain is worth it. ↩

Cite this post 
@online{swift2025fixingSlowZedAgentStartupTimes,
  author = {Ben Swift},
  title = {Fixing slow Zed agent startup times},
  url = {https://benswift.me/blog/2025/12/17/fixing-slow-zed-agent-startup-times/},
  year = {2025},
  month = {12},
  note = {AT-URI: at://did:plc:tevykrhi4kibtsipzci76d76/site.standard.document/2025-12-17-fixing-slow-zed-agent-startup-times},
}