Using MCP tools

warpdrv ships with a built-in default MCP server — warpMCP — that provides the tools listed below. You can also extend it by adding external MCP servers through the mcp.json on the MCP page.

Built-in tools

File system

Tool	Description
file_read	Read the contents of a file, optionally a range of lines.
file_write	Write content to a file, creating or overwriting it.
file_patch	Replace a specific text segment in a file.
dir_list	List a directory’s contents with optional glob filtering and recursion.

Coding / code search

Tool	Description
rg	Fast regex search across file contents using ripgrep.
code_graph_search	Search the code graph for symbols by name, kind, or signature.
code_graph_symbol	Get detailed info about a specific symbol (type, location, source).
code_graph_list	List all symbols in a file or directory.
code_graph_callers	Find all symbols that call or reference a given symbol.
code_graph_callees	Find all symbols that a given symbol calls or references.
code_graph_ingest	Index a project’s code into the code graph.
code_graph_clear	Clear the code graph index for a project.

Shell

Tool	Description
shell_exec	Execute a shell command (bash on Linux/mac, PowerShell on Windows).

Web

Tool	Description
fetch	Perform an HTTP request and return the response.

Subagent control

Tool	Description
create_subthread	Spawn a subagent (child thread) to work on a task in parallel.
subthread_send_message	Send a message to a subagent thread and wait for its response.
superthread_send_message	Send a message back up to the parent thread.
list_subthreads	List all subagent threads of the current thread.
set_current_status	Update the status line shown for the current thread.

Chat / memory search

Tool	Description
chat_search	Search across your conversations for relevant messages.
chat_get_message	Fetch the full content of a specific chat message.
embedding_search	Semantic search over embedded content in a workspace.

Task

Tool	Description
todo	Manage a per-thread todo list (read, write, update, clear).

Using external MCP servers

You can extend the built-in tools by adding external MCP servers. These are configured in mcp.json, located at ~/.config/warpcore/mcp.json (editable from the MCP page).

The format follows the standard MCP structure — a mcpServers object where each entry is a named server with a command to launch and its args:

{
  "mcpServers": {
    "desktop-commander": {
      "command": "npx",
      "args": ["-y", "@wonderwhy-er/desktop-commander"]
    }
  }
}

In this example, desktop-commander is a filesystem and desktop automation server that runs via npx with no install step. warpdrv spawns it over stdio and merges its tools into the available set.

Once added, external tools appear in the same permission system as the built-in ones (default ask). The same warning applies: any external tool that writes files or executes commands should stay on ask.

There are hundreds of community MCP servers available — search online for one that fits your use case and add it to your mcp.json.

Permissions

Every tool call goes through a permission system with three modes: ask, allow, and deny.

Global permissions (MCP page)

On the MCP page you can set a global permission mode for every tool registered by every MCP server, including the built-in warpMCP.

• Default is ask — every time a tool is called, you’re prompted to approve it.
• You can switch a tool globally to allow (auto-approve) or deny (always block).

Keep dangerous tools on “ask”

Edit tools (file_write, file_patch), the shell tool (shell_exec), and the fetch tool should never be set to auto-allow. They can modify your files, run arbitrary commands, or make network requests. Keep them on ask so you’re always prompted before they run.

Per-thread override (right panel)

On the Chat page, the right panel has a per-thread tool permission list. This overrides the global setting for that specific conversation. You can set each tool to ask, allow, or deny for the current thread only.

How a tool call is handled

When the model calls a tool:

1. The permission system resolves the effective mode (per-thread override if set, otherwise the global mode).
2. If the mode is ask, a permission interface appears showing the tool name and its arguments. You approve or deny the call.
3. If allow, the tool runs immediately.
4. If deny, the call is blocked and the model is told it was denied.

Choosing tools per chat

Near the composer (where you type your prompt) there’s a tool selector that lets you choose which tools are active for the current chat.

Note: Changing the tool selection in the middle of a chat causes the entire prompt to be reprocessed. This is because the list of available tools is sent at the very beginning of the request (per the llama.cpp specification). Set your tools at the start of the chat.

If you want different tool sets for different tasks — e.g. read-only tools for planning, and edit/shell tools for building — that’s what modes are for. See Workflow, Part 1 — Modes.

Next steps

• Workflow, Part 1 — Modes — group tools into reusable modes.
• Workflow, Part 2 — Guardrails — add safety checks around tool usage.