Give the chat more tools with MCP servers

WebF1 Pad’s AI chat can call tools. Besides the built-ins (web search, note tools), you can connect any MCP-compatible server and its tools are offered to the model during the conversation. Two kinds exist, configured in Settings → MCP Servers → Chat tools (MCP servers).

What the built-in server already gives the chat

WebF1 Pad ships with its own MCP server (the same one external agents use, see Let AI assistants work with your notes). It enables itself automatically the first time you turn on AI chat (turning it off again in Settings → MCP Servers is always respected), so the chat gains its tools with zero configuration:

• list_notes / get_note / update_note, see which notes are open, read them and edit them in place.
• list_chats / get_chat, see which chat sessions exist and read one.
• list_how_to / get_how_to, the app’s How-To documentation. Ask things like “how do I set up cloud backup?” or “which shortcuts exist?” and the model can pull the official guide instead of guessing.

The chat’s tools live behind the Tools toggle (enabled by default for tool-capable models), so the model only calls servers when it decides a tool would help. Tool call arguments are sent to the server you configured, the composer shows a chip listing which servers are active for the current chat.

The content-access gate

By default the note/chat tools can only read and update the content of the active note tab and the active chat session, anything else fails and the model must ask you to open it first. This is a setting, not a fixed rule: set Settings → MCP Servers → Content agents can read to All open notes and chats to let the AI read any open note or chat, or to Only notes marked “Always allowed” to require an explicit opt-in per note. Titles and timestamps of all open items are always visible for disambiguation; only content is gated. A per-item override always wins over the mode: right-click a note tab, or a chat session in the sidebar, → Agent access to mark it Never (the AI cannot see it at all, in any mode) or Always allowed (readable in every mode). Notes the AI created via save_note are marked “Always allowed” automatically, until you change the override. On a saved note, “Never” is durable: the file/cloud object is recorded in a local registry, so the note stays hidden from agents after you close and reopen it (even while it is the active tab) and after you move or rename the file. Only moving the choice off “Never” releases it.

Add a remote MCP server

A remote server is an HTTP endpoint speaking the MCP Streamable HTTP transport, it can run on your machine or anywhere on the internet.

1. Open Settings → MCP Servers → Chat tools (MCP servers).
2. Press Add remote server.
3. Give it a name (used as the tool namespace, e.g. a server named Gitea yields tools like mcp__gitea__list_issues), and paste its URL.
4. If the server needs authentication, add the header (typically Authorization with value Bearer <token>; the bare token without the Bearer prefix works too — the pad adds the scheme itself).
5. Press Test connection, it lists the tools the server exposes, then Save MCP settings.

The Add sample server… dropdown offers a curated set of keyless public endpoints (weather, recipes, documentation search and more) you can add with one click, plus local command samples (desktop only): Chrome DevTools, Playwright, Filesystem, Memory and Sequential Thinking. The local samples run via npx, so they need Node.js installed, see below.

Add a local MCP server (desktop)

A local server is a command that is started on this workstation and spoken to over stdin/stdout, the same way tools like npx-based MCP servers run.

1. Open Settings → MCP Servers → Chat tools (MCP servers).
2. Press Add local command (desktop only, mobile builds cannot spawn processes), or pick one of the Local samples from the Add sample server… dropdown.
3. Give it a name and type the full command line, for example npx -y some-mcp-server --flag=value. Arguments are split on spaces; quote parts that contain spaces, e.g. bash -lc "source file && exec cmd".
4. Optionally add environment variables the process should receive.
5. Press Test connection, then Save MCP settings.

Install the tools

The bundled local samples run via npx, which is part of Node.js. You install Node.js once; everything else depends on which sample you picked.

Step 1, install Node.js (required for every local sample):

• Windows: download the LTS installer (.msi) from nodejs.org and run it. Keep the default options, they add Node to your PATH.
• macOS: run brew install node (Homebrew), or use the LTS installer (.pkg) from nodejs.org.
• Linux: install from your distro’s repository, e.g. sudo apt install nodejs npm (Debian/Ubuntu), sudo dnf install nodejs npm (Fedora), or sudo pacman -S nodejs npm (Arch). For a newer version, use nvm or NodeSource’s setup script instead.

Verify in a terminal: node -v and npx -v should both print a version. If the Test button instead reports “failed to spawn npx”, Node is missing or not on your PATH (on Windows, sign out/in once after installing if the terminal still can’t find it).

Step 2, per-sample setup:

• Chrome DevTools, install Google Chrome from google.com/chrome (it drives your local Chrome, so the stable channel is enough). No further commands needed; see chrome-devtools-mcp for details.
• Playwright, Node.js is enough to start, but a browser must be downloaded once: open a terminal and run npx playwright install chromium (about 150 MB). See Playwright MCP.
• Filesystem, nothing to install, but the sample command contains a <directory> placeholder. Edit the command after adding it and replace <directory> with the real folder you want to expose, e.g. npx -y @modelcontextprotocol/server-filesystem /home/you/Documents. Careful: the server gets access to every file in that directory and all of its subdirectories — point it at a small scratch directory meant for this, not a broad location like your home folder or Documents.
• Memory and Sequential Thinking, no additional setup; they run with Node.js alone.

First run is slow: npx -y downloads the server’s package the first time it is used, so the first Test or tool call can take tens of seconds, later runs start instantly.

The process starts when the chat first uses one of its tools and stops afterwards. Local servers run with your user privileges, only add commands you trust. An unreachable or failing server never breaks the conversation: the model just receives an error result for that call.

Tool calling must be enabled

MCP tools are only offered to tool-capable models. In Settings → AI Chat, tick Tools on the provider (or on a specific model to override), the same switch that enables web search and the note tools. During a chat you can flip the Tools chip off to answer with the model’s own knowledge only.