Editora 0.20.0 docs. You are reading an archived version, kept so links from that release keep working. Read the 0.21.0 docs →

Beta MCP server

Editora can run a small Model Context Protocol server inside the editor, so an LLM agent (Claude Code, for example) can see what you’re working on and act through Editora’s own commands. It’s off by default and guarded by a security-notice dialog; enable it in Settings → MCP Server.

What it exposes

A loopback-only HTTP/JSON-RPC server with bearer-token auth exposes fourteen tools, in three groups:

Group	Tools
Reads	`list_open_files`, `list_tabs`, `read_buffer`, `get_selection`, `get_diagnostics`, `document_symbols`, `git_status`, `todo_scan`, `find_in_files`, `list_commands`
Writes	`edit_buffer` (undoable str-replace edits), `save_buffer`
Actions	`open_file`, `execute_command`

So an agent can observe live state, make undoable edits, and drive the editor through the same command registry the palette uses. It runs on the JDK’s built-in HTTP server, so there’s no new dependency.

Editing and saving

edit_buffer replaces old_text with new_text in an open buffer, through the editor’s undo history. Nothing is written to disk until save_buffer.

old_text must be the exact text to replace and must occur once in the buffer, unless replace_all is true.
To replace the buffer’s entire text, the agent passes replace_whole_buffer: true and no old_text. An edit with a missing or empty old_text is rejected, never widened to the whole buffer. A whole-buffer replacement is also refused when the buffer changed since the agent last read it with read_buffer; it reads the buffer again and retries.

save_buffer waits for the write. Its result is saved only when the file is on disk, and an error when the save was refused, failed, or is still pending. execute_command reports only that a command ran, so an agent that needs to know a save reached the disk uses save_buffer rather than running file.save.

Argument checks

Every tool validates its arguments before doing anything:

An argument the tool doesn’t define is rejected, with the accepted names in the error, so a misspelled argument can’t be silently ignored.
A path must be absolute. An empty or relative path is rejected. Tools that take an optional path use the active buffer when it is left out.

Enabling and connecting

Turn it on in Settings → MCP Server (or run View: Toggle MCP Server, view.toggleMcp), and accept the security notice.
A status-bar MCP indicator shows while it’s running. Click it, or run MCP: Copy Endpoint Command (mcp.copyEndpoint), to copy a ready-to-paste claude mcp add command with the endpoint URL and token.
The endpoint is also written to mcp-endpoint.json in your config folder for discovery.

Security

The server binds to loopback only (never the network), requires the bearer token, stays off until you enable it, and shows a security notice first. Even so, execute_command can run any command, so only connect agents you trust. Simple UI mode turns it off.

mcp-endpoint.json holds the endpoint URL and the bearer token. Where the filesystem supports POSIX permissions (so not on Windows), Editora restricts the file to your user account. It deletes the file when the server stops and leaves it out of Export Configuration. The token is the only thing that stops another program on your computer from driving the editor, so any program that can read that file can connect. The port and token are new each time the server starts: after restarting Editora or toggling the server, copy the connection command again.

Commands: view.toggleMcp, mcp.copyEndpoint.