Skip to main content

​
The /doctor command

Run /doctor inside a Claude Code session to automatically diagnose your environment. The command checks your installation, authentication status, network connectivity, MCP server connections, and configuration files, then reports any problems it finds with suggested fixes. 
/doctor

​
Common issues

Symptoms: Claude Code prompts you to log in repeatedly, or API calls fail with authentication_failed.Fixes:
  1. Re-run the login flow: 
    /login
  2. Check that your API key or OAuth token is set correctly. If you use an API key, verify the environment variable: 
    echo $ANTHROPIC_API_KEY
  3. If you are behind a corporate proxy or VPN, ensure that api.anthropic.com is reachable: 
    curl -I https://api.anthropic.com
  4. On macOS, if the keychain lookup is failing, try logging out and back in: 
    /logout
/login
  5. For third-party providers (Bedrock, Vertex, Foundry), check that the respective credentials are configured in your environment — Claude Code uses AWS credentials, gcloud ADC, or Foundry tokens depending on the provider.
Symptoms: Claude Code asks for permission repeatedly, or a tool call fails with a permission error.Fixes:
  1. Check the current permission mode with /config. Set it to acceptEdits to auto-approve file operations, or bypassPermissions in a trusted sandbox.
  2. Review your allowedTools and deniedTools settings in ~/.claude/settings.json or .claude/settings.json.
  3. If running in headless mode (-p), use --allowedTools to explicitly allow the tools Claude needs: 
    claude -p "Run tests" --allowedTools "Bash(npm test)" "Read"
  4. Use --permission-mode bypassPermissions only in sandboxed environments with no internet access: 
    claude -p "Refactor code" --permission-mode bypassPermissions
bypassPermissions disables all safety checks. Only use it in fully isolated environments.
Symptoms: Claude Code reports the context window is full, responses become truncated, or you see a “context limit” error.Fixes:
  1. Run /compact to compress the conversation context. Claude summarises the conversation so far and continues with the summary in the context window.
  2. Start a fresh session with /clear if the current conversation is no longer needed.
  3. Use /cost to monitor your context window utilisation: 
    /cost
  4. For very large codebases, use --allowedTools in headless mode to restrict tools and reduce tool-call verbosity in the transcript.
  5. Consider enabling fast mode from settings — it uses a smaller, faster model for sub-tasks, which consumes fewer tokens.
Symptoms: MCP tools are missing, or Claude Code reports an MCP server as failed or pending.Fixes:
  1. Run /doctor — it checks all configured MCP servers and reports connection errors.
  2. Run /mcp to see the connection status of each server.
  3. Check your MCP configuration file for syntax errors: 
    # For project-level MCP config:
cat .mcp.json | python3 -m json.tool

# For user-level MCP config in settings:
cat ~/.claude/settings.json | python3 -m json.tool
  4. Enable debug output to see MCP server stderr: 
    claude --debug
  5. Test the MCP server command directly in your terminal to confirm it starts without errors: 
    node /path/to/mcp-server.js
  6. If the server requires environment variables, add them to the env field in the server config: 
    {
  "myserver": {
    "type": "stdio",
    "command": "node",
    "args": ["server.js"],
    "env": { "DATABASE_URL": "postgres://..." }
  }
}
  7. In headless mode, use --mcp-config to load a config file explicitly, and --strict-mcp-config to ignore all other MCP configuration: 
    claude -p "Query the database" --mcp-config ./mcp.json --strict-mcp-config
Symptoms: API costs are higher than expected, or rate limits are hit frequently.Fixes:
  1. Use /cost to see a breakdown of token usage for the current session: 
    /cost
  2. Enable fast mode in settings. Fast mode uses a smaller model for routine sub-tasks and significantly reduces cost per session.
  3. Use /compact regularly to compress the context window rather than letting it fill up.
  4. In headless mode, use --max-budget-usd to cap spending per run: 
    claude -p "Large task" --max-budget-usd 0.25
  5. Use --max-turns to limit agentic turns in scripts: 
    claude -p "Fix lint errors" --max-turns 10
  6. Restrict available tools with --allowedTools — fewer tools means fewer tool-call round-trips: 
    claude -p "Review code" --allowedTools "Read,Glob"
Symptoms: Running claude in the terminal gives command not found.Fixes:
  1. Check that npm’s global bin directory is in your PATH: 
    npm bin -g
# Add the output to your PATH in ~/.bashrc or ~/.zshrc
export PATH="$(npm bin -g):$PATH"
  2. Reload your shell profile: 
    source ~/.bashrc   # or ~/.zshrc
  3. If you used npx to install, try a global install instead: 
    npm install -g @anthropic-ai/claude-code
  4. On macOS with Homebrew-managed Node, the global bin path may differ. Find it with: 
    which node
# e.g. /opt/homebrew/bin/node
# → global bin is /opt/homebrew/bin
  5. Verify the binary exists after install: 
    ls "$(npm bin -g)/claude"

​
Log locations

Claude Code writes logs to the following locations:
LogLocation
Session transcripts~/.claude/projects/<project-hash>/<session-id>.jsonl
Debug logsWritten to stderr when --debug is passed, or to a file with --debug-file <path>
MCP server outputShown in stderr when --debug is passed
To write debug logs to a file for analysis: 
claude --debug-file /tmp/claude-debug.log

​
Resetting configuration

If your Claude Code configuration becomes corrupted or you want a clean slate, you can reset it by clearing the ~/.claude/ directory: 
# Back up first
cp -r ~/.claude ~/.claude.bak

# Remove all config, sessions, and cached data
rm -rf ~/.claude
This removes all sessions, settings, and stored permissions. Back up anything you want to keep before proceeding.
To reset only settings without losing session transcripts: 
rm ~/.claude/settings.json
rm ~/.claude/settings.local.json

​
Checking token usage with /cost

The /cost command displays token usage and estimated cost for the current session: 
/cost
It shows:
  • Input tokens used (including cache reads and cache creations)
  • Output tokens generated
  • Total estimated cost in USD
  • Context window utilisation

​
Reporting issues

If you encounter a bug or unexpected behaviour, report it on GitHub: github.com/instructkr/claude-code When filing an issue, include:
  • The claude --version output
  • Your operating system and Node.js version
  • The command you ran (redact any secrets)
  • The full error output (use --debug-file to capture it)
  • Whether the issue is reproducible in a fresh directory

Build docs developers (and LLMs) love

Get started for freeTalk to us