Hashfile MCP Server

Fine-Grained Control for AI File Operations — A Model Context Protocol (MCP) server that gives you surgical precision over file editing and ironclad control over file access.

Why Hashfile MCP?

🎯 Surgical File Editing

Unlike traditional line-number-based editing that breaks when files change, Hashfile uses content-anchored operations:

• Hash-Anchored Lines: Every line tagged with a content hash — edits target the right content, not just the right line number
• Fuzzy Matching: Automatically finds moved lines even after insertions/deletions
• Collision Detection: File-level verification prevents editing stale content
• Multi-Operation Edits: Apply multiple precise changes in a single atomic operation

🔒 Ironclad Access Control

Take complete control over what AI agents can touch with AGENTS.md frontmatter:

• forbidden: Block access to secrets, credentials, sensitive data
• read_only: Allow reading schemas, configs, lock files — prevent modifications
• ignore: Hide generated code, dependencies, build artifacts from AI context
• Hierarchical Discovery: Place AGENTS.md anywhere in your tree — nearest file wins
• Glob Patterns: Fine-tune access with wildcards (secrets/**, **/*.env)

📖 Full AGENTS.md Documentation | 📋 Proposal Spec

🛠️ Drop-In Filesystem Compatibility

All standard MCP filesystem tools included:

• list_directory — Browse with [FILE]/[DIR] prefixes
• directory_tree — Compact tree view (10x more token-efficient than JSON)
• create_directory — Recursive directory creation
• move_file — Rename/move files and directories
• write_file — Raw UTF-8 writes (non-hashline)
• read_multiple_files — Batch reads in one operation

Features at a Glance

Installation

cargo build --release

The binary will be at target/release/hashfile-mcp.

Quick Start

1. Configure Your MCP Client

Add to Claude Desktop or your MCP client config:

{
  "mcpServers": {
    "hashfile": {
      "command": "/path/to/hashfile-mcp/target/release/hashfile-mcp"
    }
  }
}

2. (Optional) Add Access Control

Create AGENTS.md in your project root:

---
forbidden:
  - "secrets/**"
  - "**/*.env"
  - ".git/**"

read_only:
  - "package-lock.json"
  - "Cargo.lock"
  - "schema.sql"

ignore:
  - "node_modules/**"
  - "target/**"
  - "**/*.generated.ts"
---

# Project Instructions

Your custom instructions for AI agents here...

3. Start Editing

The AI can now:

• ✅ Read and edit source files with hash-anchored precision
• ✅ Browse directories and create new files
• ❌ Cannot touch your secrets or .env files
• ❌ Cannot modify lock files or schemas
• 🙈 Won't see node_modules or build artifacts

How Hashline Works

The Problem

Traditional line-number editing fails when files change:

Agent: "Replace line 42"
Reality: Someone inserted 3 lines at the top
Result: Wrong line replaced! 💥

The Solution

Content-anchored editing with hash verification:

1. Read: Each line gets a hash tag → `42:a

... View full README on GitHub