Reproduction

1. Install a MCP Server Extension (in my case, mcp-server-github)
2. Go to Zed's AI Agent window
3. Make sure you are using a profile in which a MCP server extension is enabled (in my case, write)
4. Send a message requesting to use a MCP server tool (in my case, mcp-server-github/issue_read, using the prompt : "From MCP Server mcp-server-github, use the tool issue_read to read https://github.com/zed-industries/zed/issues/47111, and describe what you see")
5. Assess that the MCP tool answer is visible when expanding them
   
6. Close Zed
7. Open Zed again
8. Go back to the original thread
9. Assess that the MCP tool answer is now broken
   

Behavior

Current

AI Agent tool call answers are cleared upon app close

However, I'm not sure if new messages sent after zed has been reopened have access to the MCP tool call answer (I'm guessing not), If so, I also doubt that token count is being recalculated (needs confirmation from someone who knows the internals)

Expected

AI Agent tool call answers remains upon app close

Also making messages sent after zed has been reopened include those tool call answers

Zed version and system specs

Zed: v0.217.4+stable (Zed)
OS: Linux Wayland nixos 25.11
Memory: 13.5 GiB
Architecture: x86_64
GPU: AMD Radeon Graphics (RADV RENOIR) || radv || Mesa 25.2.6

Related

• AI Agent: edits are implicitly kept upon app close #47528