mcp-language-server
- Type
- MCP server
- Scanned
- Nov 12, 2026
- Version
- e4395849a52e
- Files read
- 278
This MCP server runs a local Language Server Protocol (LSP) executable such as gopls, Pyright, or rust-analyzer to give AI coding agents semantic navigation, diagnostics, symbol renaming, and line-based file editing on a project workspace. Its biggest risk is that none of its file tools enforce the configured --workspace boundary, so a prompt-injected agent can overwrite arbitrary existing files (such as shell startup scripts or .git/config) via edit_file or read sensitive files outside the project via hover.
What it can reach
- Files
- Reads, edits, renames, creates, and deletes files in the configured workspace directory, and because workspace boundaries are not enforced, can also read and overwrite existing files anywhere on the user's filesystem.
- Shell commands
- Spawns and communicates over stdio with the local language server executable specified by the --lsp flag and passes any arguments after --.
- Secrets
- Passes all environment variables from the MCP server process to the child language server process, and writes logs to LOG_FILE if set.
Findings
-
HighUnrestricted file writes outside the workspace directory
internal/tools/edit_file.go:22
Even though the server requires a --workspace directory flag at startup, edit_file (and the underlying utilities.ApplyWorkspaceEdit helper used by edit_file, rename_symbol, and LSP workspace/applyEdit requests) never checks that target file paths stay inside that workspace or outside .git. If prompt injection in source code, an issue, or a webpage tricks the AI agent into calling edit_file on a path such as ~/.zshrc, ~/.bashrc, an agent config file, or .git/config (or on a symlink inside the workspace pointing elsewhere), the server overwrites that file on disk, which can lead to arbitrary code execution.
Fix: Resolve target paths with filepath.Abs and filepath.EvalSymlinks in both the tool handlers and utilities.ApplyWorkspaceEdit, verify they stay strictly inside the configured workspace directory, and block modifications inside .git.
-
HighArbitrary file reading outside the workspace via hover and diagnostics
Neither hover nor diagnostics checks that filePath is within the configured --workspace directory before reading the file with client.OpenFile and os.ReadFile. In GetHoverInfo, when the language server returns empty hover content for a position (which occurs on plain-text and secret files with language servers such as Pyright, TypeScript, and clangd), the tool falls back to ExtractTextFromLocation and returns the raw line of text from disk, allowing a prompt-injected AI agent to read sensitive files like ~/.ssh/id_rsa, ~/.aws/credentials, or .env files line by line.
Fix: Validate in client.OpenFile, GetHoverInfo, GetDiagnosticsForFile, and ExtractTextFromLocation that the symlink-resolved file path resides inside the configured workspace directory before reading from disk.
