Claude Code MCP Server Setup: PandaNpc

Claude Code MCP server setup connects your agent to PandaNpc notes, model configurations, and scheduled tasks. Register the server, sign in, and verify its tools.

Updated

The PandaNpc MCP server connects an MCP-compatible coding agent to your PandaNote knowledge base, work logs, scheduled tasks, model configuration, and Agent Service Center. It is useful when an agent needs durable project context instead of starting from an empty conversation every time.

The server runs locally over MCP's standard input/output transport. Your MCP client starts the pandanpc-mcp process when a session opens; you do not need to expose a port or run a separate web server.

Before you install

You need:

  • Node.js and npm available in the same environment where the MCP client runs.
  • A PandaNpc account with access to the email address used for verification.
  • Claude Code, OpenAI Codex, Cursor, or another client that supports local stdio MCP servers.

If you want the agent to schedule work on one of your own computers, install and connect PandaPaw first. The notes and workspace tools do not require an online PandaPaw device.

Install or update the MCP server

Install the package globally on macOS, Linux, or Windows:

bash
npm install -g @pandanpc/mcp-server

Confirm that the executable is visible to your shell:

bash
pandanpc-mcp --help

Run the same npm command whenever you want to update. Restart the MCP client after an update because an already-running client keeps the process version it originally launched.

Check the MCP server version history when a tool or its behavior differs from an older screenshot or tutorial.

Add PandaNpc to Claude Code

The most reliable method is Claude Code's own configuration command. Use user scope if you want PandaNpc available in every project:

bash
claude mcp add --scope user pandanpc -- pandanpc-mcp

Use --scope project instead when the configuration should be shared through the current repository, or --scope local when it should apply only to you in the current project.

Check the registered server:

bash
claude mcp get pandanpc
claude mcp list

You can also configure the stdio server manually in the MCP configuration used by your Claude Code installation:

json
{
  "mcpServers": {
    "pandanpc": {
      "command": "pandanpc-mcp"
    }
  }
}

Add PandaNpc to OpenAI Codex

Codex can register the same local process from the command line:

bash
codex mcp add pandanpc -- pandanpc-mcp

Verify the saved configuration:

bash
codex mcp get pandanpc
codex mcp list

The equivalent entry in ~/.codex/config.toml is:

toml
[mcp_servers.pandanpc]
command = "pandanpc-mcp"

Restart Codex or open a new session after adding the server. If you maintain a custom Codex configuration file, put the MCP table in that file rather than duplicating it across files.

Configure Cursor and other MCP clients

For Cursor, add a local server to ~/.cursor/mcp.json using the same JSON shape shown for Claude Code. For Windsurf and other clients, create a stdio MCP server whose command is pandanpc-mcp and whose argument list is empty.

The executable must be on the PATH inherited by the graphical application. If it works in a terminal but the desktop client reports “command not found,” use the absolute executable path returned by the following command:

bash
command -v pandanpc-mcp

On Windows PowerShell, use:

powershell
Get-Command pandanpc-mcp

Sign in from the agent

PandaNpc uses a two-step email verification flow. Ask the agent to perform these actions in order:

  1. Call note_login with your email address. PandaNpc sends a verification code.
  2. Call note_login again with the same email address and the code.

The local credential is cached, so you should not need to sign in during every session. Never paste the verification code into a public transcript or a shared prompt.

Verify the MCP connection

Run a small read-only test before allowing the agent to create or edit data:

  1. Ask it to call list_workspaces.
  2. Ask it to call list_notes for one workspace.
  3. Ask it to search for a harmless keyword with search_notes.

A successful response proves that the MCP client started the process, authentication completed, and the account API is reachable. For a write test, create a disposable note and then soft-delete it.

The exact tool list shown by your MCP client is the source of truth for the installed version. Tools are grouped around these workflows:

Workflow Typical tools Important behavior
Notes list_notes, create_note, read_note, update_note, search_notes update_note replaces the complete body; read before editing
Work logs save_done Stores completed work in today's project folder automatically
Workspaces list_workspaces, create_workspace, update_workspace A bound project path lets the server find the correct workspace
Scheduled work list_scheduled_bridges, create_scheduled_task, list_scheduled_task_runs A PandaPaw machine must be online; cron expressions use six fields
Providers and models list_providers, list_models, create/update/delete tools List first so you use current IDs and valid model types
Rules and skills sync_rules Synchronizes supported project instructions while skipping unchanged files

For provider and model field definitions, see Custom AI providers and API keys. For account and billing boundaries, see PandaNpc accounts, plans, and API costs.

Safe operating practices

  • Start with list, read, or search tools before create, update, delete, or schedule actions.
  • Read a note before updating it because note updates replace the whole Markdown document rather than applying a patch.
  • Treat scheduled prompts as code execution requests: confirm the target machine, working directory, timezone, and six-field cron expression.
  • Keep provider API keys out of prompts and notes. Enter secrets only through the intended provider configuration flow.
  • Review the MCP tool approval card before allowing destructive or external actions.
  • Use save_done for work logs instead of manually guessing the correct workspace and date folder.

Troubleshooting

The client says `pandanpc-mcp` was not found

Open a new terminal after installation and check the executable path. Desktop applications sometimes inherit an older PATH; configure the absolute path or restart the application. If npm was installed through a version manager, make sure the MCP client starts under the same user and environment.

The server is registered but no tools appear

Restart the client, then inspect claude mcp get pandanpc or codex mcp get pandanpc. Remove and add the entry again if the command is wrong. Also check that you did not configure the local stdio process as an HTTP URL.

Tools say you are not authenticated

Repeat the two note_login calls. Use the newest verification code, check the spam folder, and confirm that both calls use the same email address.

Scheduled tasks cannot find a machine

Call list_scheduled_bridges first. If it returns no available device, verify PandaPaw status and device connectivity on the target machine. A machine that is powered off, sleeping, signed out, or disconnected cannot start a scheduled run.

A note update removed existing text

update_note accepts the full replacement body. Restore the note from its history or trash where available, then read the current content and submit the combined document.

Need account-specific help? Contact support@pandanpc.com and include the MCP client name, operating system, package version, and the exact error message without secrets.