Cursor CLI is getting slower and slower as chat and token size increase
When using MCP tools with the cursor cli, it is getting quite slow (taking 3-4 mins just to analyze the next move)
Steps to Reproduce
Start chat with cursor-agent
Approve mcp server
Provide the prompt to perform actions that require the MCP tools
After a few tool calls, it will get slow and keep getting slower with time
Expected Behavior
It should work as seamlessly as it does in the cursor IDE
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
Hey, thanks for the report. Could you please provide:
Which MCP server are you using? (e.g., filesystem, git, custom)
Approximate conversation size when the slowdown starts? (number of messages/tool calls)
Does restarting cursor-agent temporarily help?
Request ID: please share the request ID from the problematic chat
In the meantime, you could try:
Starting a new chat when it begins to slow down
Using fewer tool calls per session if possible
If this is reproducible with specific steps, I’ll escalate it to the engineering team, since it looks like the CLI might not be managing context/memory as efficiently as the IDE.
Conversation is getting slower after 2-3 tool calls itself, and around 10k token size, and then it keeps getting slower
Restarting the agent is not helping. The same issue persists.
I am getting ‘no rquestid found‘ error while hitting ‘copy-request-id’ command from CLI
We can not start a new chat after a few tool calls, as we need chat context
/history within the session and when the session ends. Also, we will need multiple and repetitive tool calls for our use case with the cursor CLI.
The issue is reproducible every time with the shared steps. The same task with tool calls takes around 5-7 minutes in the IDE and around 25 minutes in the CLI.
Looking forward to your response and hoping for a quicker resolution on this issue.
Thanks for the detailed clarification! It looks like a memory/context management issue in the CLI, which we already have on our radar.
Given that:
Performance drops sharply after 2–3 tool calls
It’s about 5× slower than in the IDE for the same operations
The copy-request-id command crashes
Unfortunately, there’s no reliable workaround if it still reproduces after a restart.
Could you share details about your custom MCP server (what it does, typical response sizes, whether streaming is enabled) to help us reproduce this faster?
The custom MCP server is created for performing actions/interactions(take screenshots, clicks, etc.) on mobile apps. It returns images and small to mid-sized (100-5k token size) responses. Streaming is not used.
As you mentioned, the memory/context management issue in the CLI is already on your radar. Can you please share by when we can expect the resolution on the same?
Thanks for the additional information about your MCP server. This helps the engineering team reproduce the issue more effectively.
The CLI memory and context management issue is actively being tracked in our development backlog. Unfortunately, I can’t provide a specific timeline for a fix, but the engineering team understands how much this impacts your workflow (especially scenarios that require extended context like yours).