Every time an AI coding assistant edits a file, it normally needs two tool calls: one to Read the file, then one to Edit it. Refactoring across 5 files? That's 10 calls. Refactoring across 20? That's 40 calls — each one burning tokens on JSON overhead, waiting for round-trips, and filling up context.
MCP ReadEdit collapses those pairs into single calls. Read+Edit in one shot. Batch edits across many files in one call. The result: 80–95% fewer tool calls, faster completions, and significantly lower token usage.
Combine Read+Edit into single tool calls — 80-95% fewer tool calls for multi-file refactoring.
An MCP server that gives any AI coding assistant batch file operations. Instead of separate Read → Edit calls per file, do it all in one shot.
- Node.js 20+ (recommended: latest LTS)
No install needed — run directly with npx:
npx mcp-readeditOr install globally for faster startup:
npm install -g mcp-readeditThen add it to your MCP client (see Client Setup below).
Add to ~/.claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"readedit": {
"command": "npx",
"args": ["mcp-readedit"]
}
}
}claude mcp add readedit -- npx mcp-readeditAdd to .cursor/mcp.json in your project:
{
"mcpServers": {
"readedit": {
"command": "npx",
"args": ["mcp-readedit"]
}
}
}Go to Settings → MCP Servers and add:
{
"readedit": {
"command": "npx",
"args": ["mcp-readedit"]
}
}In Cline settings, add to MCP Servers:
{
"readedit": {
"command": "npx",
"args": ["mcp-readedit"]
}
}Add to .continue/config.yaml:
mcpServers:
- name: readedit
command: npx
args:
- mcp-readeditAdd to your Zed settings.json:
{
"context_servers": {
"readedit": {
"command": "npx",
"args": ["mcp-readedit"]
}
}
}| Tool | What it does |
|---|---|
read_edit |
Read a file, optionally edit it — 1 call instead of 2 |
multi_edit |
Edit multiple files at once |
multi_read_edit |
Read + optionally edit multiple files — the powerhouse |
get_gain |
Show your token savings statistics |
Read a file and optionally replace text in one call. Returns file content.
{
"file_path": "/absolute/path/to/file.ts",
"old_string": "text to replace",
"new_string": "replacement text"
}Options: use_regex (boolean), replace_all (boolean), offset (line number), limit (line count). Omit old_string/new_string to just read.
Batch edits across files in a single call. Use when you already have the file contents.
{
"edits": [
{ "file_path": "/path/a.ts", "old_string": "foo", "new_string": "bar" },
{ "file_path": "/path/b.ts", "old_string": "baz", "new_string": "qux", "replace_all": true }
]
}The most powerful tool. Read and optionally edit any number of files in one call.
{
"operations": [
{ "file_path": "/path/a.ts" },
{ "file_path": "/path/b.ts", "old_string": "old", "new_string": "new" },
{ "file_path": "/path/c.ts", "old_string": "\\d+", "new_string": "0", "use_regex": true }
]
}Options: include_content (boolean, default false) and include_original (boolean, default false) control what's returned.
{ "breakdown": "summary" }Breakdown types: summary (default), daily, recent, all.
Refactoring a feature across 9 files:
Without MCP ReadEdit — 52 tool calls:
Read file1 → Edit file1 → Read file2 → Edit file2 → ... → Read file9 → Edit file9
28 Edit + 19 Read + 5 Write = 52 calls
With MCP ReadEdit — 4 tool calls:
multi_read_edit (files 1-3) → multi_read_edit (files 4-6) → multi_read_edit (files 7-9) → multi_edit (final batch)
Result: 48 calls saved (~9,600 tokens)
Each tool call is recorded to a local SQLite database. The tracker calculates what it would have taken with standard Read+Edit calls:
read_editwith edit: 2 standard calls → 1 optimized callmulti_edit(N files): 2N standard calls → 1 optimized callmulti_read_edit(N files): 2N standard calls → 1 optimized call
Token savings are estimated at ~200 tokens per avoided call (JSON overhead, tool result wrapping). The database auto-creates on first use.
If installed globally (npm install -g mcp-readedit), the readedit command gives terminal access to gain stats:
readedit gain # Summary
readedit gain --daily # Day-by-day breakdown
readedit gain --recent 20 # Last 20 operations
readedit gain --all # All breakdowns
readedit gain --format json # JSON export
readedit gain --reset # Reset statisticsWorks with npx too: npx mcp-readedit starts the server, readedit gain runs the CLI.
Add these instructions to your project's CLAUDE.md, AGENTS.md, or .cursorrules to make your AI coding assistant automatically use ReadEdit tools:
## File Operations — MCP ReadEdit
When reading or editing files, ALWAYS prefer MCP ReadEdit tools over separate Read + Edit calls.
### Tool Selection
- **multi_read_edit**: Read and/or edit multiple files (most common — use for any multi-file task)
- **multi_edit**: Edit multiple files when you already have their contents
- **read_edit**: Single file read-only or read+edit
- **get_gain**: Check token savings statistics
### Rules
1. NEVER use separate Read then Edit calls when ReadEdit tools are available
2. Batch file operations: group related files into a single multi_read_edit call
3. Use `use_regex: true` for pattern-based replacements
4. Read-only operations in multi_read_edit always return file content — no need to separately read files first
5. When refactoring across multiple files, plan all edits first, then execute in one multi_read_edit callFor global usage (all projects), add to ~/.claude/AGENTS.md instead.
git clone https://github.com/abnersajr/mcp-readedit.git
cd mcp-readedit
npm install
npm testIssues and PRs welcome at github.com/abnersajr/mcp-readedit.
MIT