[BUG] Claude Code - Persistent Logout & Configuration Loss Bug Report

!Image

It just happened again, I had 5 instances running in 3 different projects. I closed one instance with Ctrl + C and then started it again. It showed me the Welcome to Claude Code message, I then closed the tab in hopes that it would not be reset again. However as you can see in the screenshot it was. I have to auth myself again and add all MCP servers as well. Luckily I have them somewhere and Claude Code can add them globally based on that, but it's getting pretty annoying. No idea why it happened this time.

edit: Might worth noting everything kept working fine in the other sessions, but from experience I know that would only last until I ended the session. I have now logged back in and restored the MCP servers.

Courtesy of Claude Code:

Debug Information for Claude Code Issue #1676

MCP Server Configuration Loss After Restart

Issue Summary

Claude Code restarted itself, logged the user out, and lost all MCP server configurations. This is the third occurrence of this issue.

Environment Details

• Claude Code Version: 1.0.17
• macOS Version: 15.5 (24F74) - Darwin 24.5.0
• Node Installation: Using Homebrew (/opt/homebrew/bin/node)
• Shell: zsh with custom Claude alias
• Date: June 8, 2025

Key Findings

1. Multiple Claude Processes

There are 5 Claude CLI processes running in different terminal sessions, plus the Claude desktop app processes:

codaveto  68760  claude (terminal session s005)
codaveto  59507  claude (terminal session s016)
codaveto  56270  claude (terminal session s010)
codaveto  48768  claude (terminal session s002)
codaveto  40885  claude (terminal session s001)

This might indicate:

• Multiple concurrent sessions could be causing conflicts
• Possible race conditions when accessing configuration files

2. Configuration File State

• ~/.claude/.mcp.json: Last modified May 27, 20:11 (old timestamp)
• ~/.claude/settings.json: Empty {}
• ~/.claude/settings.local.json: Contains MCP permissions but no server configurations

``json
{
"permissions": {
"allow": [
"Bash(claude mcp add -s user dart -e DART_TOKEN=XXX npx -y dart-mcp-server)",
"Bash(claude mcp add:*)"
]
},
"enableAllProjectMcpServers": false
}
``

• No project-specific directory found for bwai

3. Missing Logs

No log files found in standard locations:

• No Claude logs in ~/Library/Logs
• No .log files in ~/.claude/
• No apparent crash reports or error logs

4. Shell Configuration

User has a custom alias in .zshrc:

# Claude fix for env -S issue
alias claude='/opt/homebrew/bin/node /opt/homebrew/bin/claude'

Potential Issues Identified

1. Configuration Storage Inconsistency: The MCP configuration appears to be stored separately from the .mcp.json file (which has an old timestamp)
2. Runtime Storage: The actual user-scoped MCP configurations might be stored in the __store.db file or elsewhere
3. Process Synchronization: Multiple Claude processes could be causing state synchronization issues
4. No Persistent Logging: Lack of logs makes it difficult to trace what happens during the failure

Recommendations for Developers

1. Add Comprehensive Logging

Implement detailed logging for:

• MCP server configuration changes (add/remove/update)
• Authentication state changes
• Configuration file reads/writes
• Process lifecycle events
• Error conditions and stack traces

2. Investigate Storage Mechanism

• The .mcp.json file timestamp suggests configs might be stored elsewhere
• Check if __store.db is where runtime configurations are actually stored
• Ensure proper file locking when multiple processes access configuration
• Document the actual storage location and format for MCP configurations

3. Handle Multiple Processes

• Implement proper IPC or file locking to prevent configuration conflicts
• Consider using a single configuration service process
• Add mutex/semaphore protection for configuration file access
• Implement configuration change notifications across processes

4. Add Recovery Mechanism

• Backup configurations before updates
• Implement configuration validation on startup
• Add ability to restore from backup if primary config is corrupted
• Consider versioning configuration files

5. User Experience Improvements

• Add clear error messages when configuration is lost
• Provide automatic backup/restore functionality
• Show warnings when multiple Claude instances are detected
• Add a claude mcp backup and claude mcp restore command

Additional Context

This appears to be a critical issue affecting user workflow, especially given it's happened three times to the same user. The lack of error logs and the fact that configurations are completely lost (not just corrupted) suggests this might be related to:

• Authentication token expiration causing a full reset
• Process cleanup routines being too aggressive
• Storage location being cleared during some update or cleanup process

Restored MCP Servers

The following servers were restored from the user's backup:

1. TAV2 (Tavily search)
2. dart (Dart package server)
3. context7 (Upstash context)
4. mcp-server-firecrawl (Web scraping)
5. figma-developer-mcp (Figma API)
6. gdrive (Google Drive)
7. supabase (PostgreSQL database)
8. supabase-postgrest (REST API)
9. markdownify (Markdown conversion)
10. puppeteer (Browser automation)
11. local-context (Local repository scanner)
12. imagen-3-0-generate-google (Google's Imagen 3.0)

Aaaand again.. 😩

!Image

!Image

I'm experiencing the same thing on PopOS 22.04 (Ubuntu based), have been experiencing this since 1.0, roughly. In my case, every time I restart my computer claude loses its configuration, mcp, everything. I'm forced to setup everything again. Sometimes a restart is not even needed to trigger this bug. Using GNOME Terminal.

Same problem. After new login on max plan all configuration is lost and gone; empty config file.

I encountered this issue twice last week. All my configuration and MCP server setup was lost. Fortunately, I have a setup script for MCP servers, but this issue is quite frustrating. I haven’t noticed any patterns that could trigger this. I didn’t restart my MacBook either.

I just had this happen twice, back to back, losing my (global) MCP configuration twice! argh

And just now a third time: seems like the latest update (1.0.29) seems to happen each time I start up

This happened to me just now when I tried using a configuration of this form:

{
  "mcpServers": {
    "language-server": {
      "command": "/full/path/to/your/clone/mcp-language-server/mcp-language-server",
      "args": [
        "--workspace",
        "/path/to/workspace",
        "--lsp",
        "language-server-executable"
      ],
      "env": {
        "LOG_LEVEL": "DEBUG"
      }
    }
  }
}

(notice the env key, when I added that claude code crashed and overwrote my config 😭 )

I believe this is actually a duplicate of #1788, which we should have fixed in the next release.

@ashwin-ant I don't think it is. #1788 just mentions a reset of MCP config. We are also being logged out and need to reauthenticate. It happened to me AGAIN a few minutes ago. First session of the day.

claude --version
1.0.31 (Claude Code)

Agreed. Also, the problem mentioned is macOS related as far as I understand. I'm experiencing this one on Debian

This absolutely still happens to me too on Ubuntu 25.04

Can you confirm if this tends to happen after modifying the an MCP configuration, or if it's just random?

@ashwin-ant not clear who you're asking. For me it was outside of MCP configuration (I have stopped reconfiguring MCPs as I expect to lose the configuration again)

I'm trying to understand this better. ~/.claude/.mcp.json is not a supported file path for MCP config. The only place where MCP configs are stored currently are ~/.claude.json and .mcp.json in your project directory. It sounds like the issue here is that ~/.claude.json is being reset in some way. If anyone's able to confirm that, that would be very helpful here.

can confirm this issue on ubuntu - has been driving me nuts. Have to log in multiple times per day, and loose all MCP configurations each time. The reset is seemingly random for me. I have stopped utilising MCPs as a result of not wanting to set them up every two hours.

Exactly; the main Claude dotfile config gets replaced and I lose my MCP config and credentials etc, as if I had just installed for the first time.

@ashwin-ant - indeed not the ~/.claude/.mcp.json, just using that as my backup source. but good to know the location of the actual config! that was a mystery to me.

i've copied its contents and will compare / checkout the contents should it happen again. i'm pretty sure it does though if I look at my firstStartTime it says 5 days ago (see screenshot, second arrow).

i'll keep an eye on this file and will update with my findings should it happen again. if you need any other details let me know

<img width="870" alt="Image" src="https://github.com/user-attachments/assets/bbe1d391-1a3c-4855-970d-75db2f1a1370" />

Something interesting happened today. I tried to start claude after exiting my terminal (warp) accidentally and it showed my my .claude.json was corrupt (image 1). I looked it up and it showed the mess you in (image 2) at line 2340. Partially duplicated the end of the JSON somehow.

After removing the duplicated partial data it all worked fine.

I'm guessing this is a new mechanic and could be the culprit of previous issues. No idea how it happened though, I haven't touched it anywhere. Perhaps closing the terminal interrupted some kind of update to the file is all I can think of, still weird that it shows double partial data. Either way hope this helps and I'm glad I still have my config.

!Image

!Image

Hey all, thanks again for the patience here. We believe we have a fix, but we're taking some extra time to make sure we get it right.

I hope so! Just happened again for the first time. I opened a window, typed claude and it somehow launched two instances shortly after each other. I quit because I thought I saw two claude interfaces in one session. Then started it again and I had to choose the theme. I saw these corrupted files in my finder and the .claude.json is pretty much empty.

!Image

!Image

I found the source issue for me, and implemented an easy fix that seems to have solved it.

When running the local installation, it would attempt an auto update - but would install the update globally rather than locally. This would result in two conflicting versions attempting to run when running "claude"

Running migrate-installer would return "Already running from local installation. No migration needed."

However, upon inspection of .claude.json it was evident that installation method was set to "global"

i corrected it to local, removed the global installation, and no longer have the issue of being logged out every 45 minutes.

Are people still seeing this in 1.0.43? We rolled out a potential fix but please let me know if it's still happening and we can keep working on this.

Hello @ashwin-ant I've actually experienced this for the first time Today after a marathon claude-coding session yesterday. It's possible that the version update with multiple instances open may have actually caused the corrupted .claude.json file.

What would be your suggested approach for recovery? Should I start with a fresh settings file (I've re-logged into my Anthropic account and it has provided a different userID which I find somewhat concerning) or try and fix the .claude.json.corrupted.xxxxxx file?

@kylesnowschwartz - did you confirm your claude --version

@igorkofman I surely did - 1.0.43 (Claude Code) - I have it on auto-update as well. FWIW the claude.json.corrupted file was 20mb large

What happens if you rename claude.json.corrupted to claude.json and launch
claude (backup your real claude.json) - it should tell you what failed to
parse on launch.

On Mon, Jul 7, 2025 at 2:42 PM kylesnowschwartz @.***>
wrote:

kylesnowschwartz left a comment (anthropics/claude-code#1676) <https://github.com/anthropics/claude-code/issues/1676#issuecomment-3046597633> @igorkofman <https://github.com/igorkofman> I surely did - 1.0.43 (Claude Code) - I have it on auto-update as well. FWIW the claude.json.corrupted file was 20mb large — Reply to this email directly, view it on GitHub <https://github.com/anthropics/claude-code/issues/1676#issuecomment-3046597633>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AADSWQRH3HIJDRYJXKMDUR33HLSTTAVCNFSM6AAAAAB6WPLWC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANBWGU4TONRTGM> . You are receiving this because you were mentioned.Message ID: @.***>

Are people still seeing this in 1.0.43? We rolled out a potential fix but please let me know if it's still happening and we can keep working on this.

yes

<img width="1051" height="508" alt="Image" src="https://github.com/user-attachments/assets/36182e10-66c1-4d22-9328-04eccdc36726" />

just happened again.

as seen in my previous comment the JSON gets malformed, it duplicates a part of it and makes the JSON invalid.

This time exactly the same happened. After restoring one of the backups everything is fine.

I tihnk it has to do with the fact that every paste i make gets saved locally am I right. This makes the JSON grow incredibly large and then when i quit claude (with multiple or single sessions(s)) by closing my terminal it doesnt have enough time toparse and save the JSON or something? I know it happens after I close Claude at least. So probably the big file size + serialisation and deserialissation

I think if we make some kind of invalid JSON fixer it might be a good workaround for now. Everytime I checked the corrupted JSON i see there is duplicated parts of the JSON added to the end of the file. when i remove that the JSON becomes valid again

Are people still seeing this in 1.0.43? We rolled out a potential fix but please let me know if it's still happening and we can keep working on this.

Yes, nothing has changed

Are people still seeing this on 1.87 or later? It's been a couple of months since we've seen this issue reported.

I consistently lose my MCP configuration: most recently had to re-re-re-re-install the user-level github MCP on Tuesday. Not sure what version that was (I keep it autoupdated). Today I am on version 1.88

I just went to check again and I am totally lost about what's going on.

I can see from my _shell history_ that I ran:

> 786ms  3d ago claude mcp add github-mcp-server -s user -- github-mcp-server stdio

From my home directory:

bollo:~ 8.8s $ claude mcp list
No MCP servers configured. Use `claude mcp add` to add a server.
bollo:~ 0.7s $ 

So!? where did it go
?
I _distinctly recall_ it saying it had written to ~/.claude.json:

bollo:~ $ ls -l ~/.claude.json
-rw-rw-r-- 1 matthew matthew 2292681 Aug 23 09:05 /home/matthew/.claude.json

which is a gargantuan document, but poring over it with jq shows _no_ github MCP.. I _know_ that before now I've seen the MCP configuration in there.

Confusingly there's also a ~/.claude directory:

bollo:~ 50.8s $ ls -l ~/.claude
total 108
drwxrwxr-x  2 matthew matthew  4096 Jul 29 11:28 agents/
lrwxrwxrwx  1 matthew matthew    50 Apr 15 08:17 CLAUDE.md -> ../.homesick/repos/dotfiles/home/.claude/CLAUDE.md
drwxrwxr-x  2 matthew matthew  4096 Jul 30 15:04 config/
drwxrwxr-x  2 matthew matthew  4096 Aug 22 09:58 ide/
drwxrwxr-x  3 matthew matthew  4096 Aug 20 09:37 plugins/
drwxrwxr-x 22 matthew matthew  4096 Aug 22 17:39 projects/
-rw-rw-r--  1 matthew matthew   726 Aug 13 11:51 settings.json
-rw-rw-r--  1 matthew matthew   145 Jun 10 15:07 settings.local.json
drwxrwxr-x  2 matthew matthew  4096 Aug 22 17:39 shell-snapshots/
drwxrwxr-x  2 matthew matthew  4096 Jun 21 12:57 statsig/
drwxrwxr-x  2 matthew matthew 69632 Aug 22 17:39 todos/

The settings.json contains:

bollo:~ $ jq -C . ~/.claude/settings.json
{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "feedbackSurveyState": {
    "lastShownTime": 1754004393588
  },
  "model": "opusplan",
  "statusLine": {
    "type": "command",
    "command": "input=$(cat); cwd=$(echo \"$input\" | jq -r '.workspace.current_dir'); hostname_short=$(hostname -s); user=$(whoami); if [ \"$user\" = \"root\" ] || [ \"$user\" = \"toor\" ]; then suffix='#'; else suffix='$'; fi; git_info=\"\"; if [ -d \"$cwd/.git\" ]; then cd \"$cwd\" && git_branch=$(git branch --show-current 2>/dev/null) && [ -n \"$git_branch\" ] && git_info=\" ($git_branch)\"; fi; printf \"\\033[2;36m%s\\033[0m:%s%s %s \" \"$hostname_short\" \"$(basename \"$cwd\")\" \"$git_info\" \"$suffix\""
  }
}

and

bollo:~ $ jq -C . ~/.claude/settings.local.json
{
  "permissions": {
    "allow": [
      "WebFetch(domain:cppcast.com)",
      "Bash(curl:*)"
    ]
  },
  "enableAllProjectMcpServers": false
}

... so where the heck did github MCP go? I haven't removed it. It has vanished.

I _know_ that I used it after re-adding it 3 days ago as it's part of a process where my CLAUDE instructions tell it to _only use the MCP_.

---

Now trying to -re-add it, to see what happens.

bollo:~/d/c/compiler-explorer (main|✔) 1.0s $ claude mcp add github-mcp-server -s user -- github-mcp-server stdio
Added stdio MCP server github-mcp-server with command: github-mcp-server stdio to user config
File modified: /home/matthew/.claude.json

so, yes, it wants to edit my ~/.claude.json as I recalled. and now there's an MCP servers section:

  "mcpServers": {
    "github-mcp-server": {
      "type": "stdio",
      "command": "github-mcp-server",
      "args": [
        "stdio"
      ],
      "env": {}
    }
  },

Just observationally there is a _TON_ of chat history context in that file. How do multiple concurrent claude sessions handle updating that file? I'll likely have 3-5 claude sessions open at once and my expectation is that adding an MCP "user level" should be "sticky" EVEN if there are running claude sessions. Is that part of it?

That said I'm sure I've also tried adding MCPs other ways.

Additionally:

• _please_ don't mix config and history in a file like this. It seems really bad and prone to this kind of problem. I'm sure if MCPs were stored in ~/.claude/mcp.json we would not have this issue
• why on earth is there a ~/.claude.json and a ~/.claude directory? Please put everything in one place.

I'm on Linux, claude doctor says:

bollo:~/d/c/compiler-explorer (main|✔) 2.7m $ claude doctor

 Diagnostics
 └ Currently running: native (1.0.88)
 └ Path: /home/matthew/.local/bin/claude
 └ Invoked: /home/matthew/.local/share/claude/versions/1.0.88
 └ Config install method: native
 └ Auto-updates enabled: true
 └ Search: OK (bundled)

 Press Enter to continue…

@dhollman

To the extent it's helpful I can share my entire .claude.json or whatever with daisy directly (and/or screenshare). Let me know :)

hey all, we thought we had landed a fix for this a few weeks ago, but evidently @mattgodbolt is stilling getting errors, even if the number of users reporting this has decreased - i'll pick this up and investigate a fix

Thank you! I haven't tried recently making any changes to configuration, but still re add my MCPs and see if I can get it to trigger again.

Been experiencing these issues as well, as well as Claude Code that keeps expiring my login sessions. I've found the VSCode extension extremely buggy as well but this is in Cursor which I assume isn't tested.

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.