Transform any MCP server into executable, type-safe TypeScript tools using progressive loading pattern. Achieve 98% token savings by loading only what you need.
Tip
Inspired by Anthropic's engineering blog post on Code Execution with MCP.
Traditional MCP integration loads ALL tools from a server (~30,000 tokens), even when you only need one or two. This wastes context window space and slows down AI agents.
Progressive loading generates one TypeScript file per tool (~500-1,500 tokens each). AI agents discover and load only what they need via simple ls and cat commands.
Result: 98% token savings + autonomous execution + type safety
cargo install mcp-execution-cliDownload from GitHub Releases:
# macOS (Apple Silicon)
curl -L https://github.com/bug-ops/mcp-execution/releases/latest/download/mcp-execution-cli-macos-arm64.tar.gz | tar xz
# macOS (Intel)
curl -L https://github.com/bug-ops/mcp-execution/releases/latest/download/mcp-execution-cli-macos-amd64.tar.gz | tar xz
# Linux (x86_64)
curl -L https://github.com/bug-ops/mcp-execution/releases/latest/download/mcp-execution-cli-linux-amd64.tar.gz | tar xz
# Linux (ARM64)
curl -L https://github.com/bug-ops/mcp-execution/releases/latest/download/mcp-execution-cli-linux-arm64.tar.gz | tar xz
# Windows (x86_64)
# Download mcp-execution-cli-windows-amd64.zip from releases pagegit clone https://github.com/bug-ops/mcp-execution
cd mcp-execution
cargo install --path crates/mcp-cliAdd individual crates to your project:
cargo add mcp-execution-core # Foundation types
cargo add mcp-execution-introspector # MCP server analysis
cargo add mcp-execution-codegen # TypeScript generation
cargo add mcp-execution-files # Virtual filesystem
cargo add mcp-execution-skill # SKILL.md generationOr add to Cargo.toml:
[dependencies]
mcp-execution-core = "0.7"
mcp-execution-introspector = "0.7"
mcp-execution-codegen = "0.7"Important
Requires Rust 1.91 or later for building from source.
# 1. Configure MCP server in ~/.claude/mcp.json
# 2. Generate tools
mcp-execution-cli generate --from-config github
# Output: ~/.claude/servers/github/
# - createIssue.ts
# - updateIssue.ts
# - ... (one file per tool)# List available tools
ls ~/.claude/servers/github/
# Load only what you need (98% token savings!)
cat ~/.claude/servers/github/createIssue.ts
# Execute autonomously
node ~/.claude/servers/github/createIssue.ts --repo="owner/repo" --title="Bug"Note
Each tool file is self-contained with full TypeScript interfaces and JSDoc documentation.
| Feature | Description |
|---|---|
| 98% Token Savings | Load 1 tool (~500 tokens) vs all tools (~30,000 tokens) |
| Autonomous Execution | Generated files run directly via Node.js CLI |
| Type-Safe | Full TypeScript interfaces from MCP JSON schemas |
| Lightning Fast | 526x faster than target (0.19ms for 10 tools) |
| 100% MCP Compatible | Works with all MCP servers via rmcp SDK |
| Thoroughly Tested | 626 tests with 100% pass rate |
| Crate | Description |
|---|---|
| mcp-execution-core | Foundation types, traits, and error handling |
| mcp-execution-introspector | MCP server analysis using rmcp SDK |
| mcp-execution-codegen | TypeScript code generation with progressive loading |
| mcp-execution-files | Virtual filesystem for code organization |
| mcp-execution-skill | SKILL.md generation for Claude Code |
| mcp-execution-server | MCP server for progressive loading generation |
| mcp-execution-cli | Command-line interface |
Dependency Graph:
mcp-execution-cli → {mcp-execution-codegen, mcp-execution-introspector, mcp-execution-files, mcp-execution-core, mcp-execution-skill}
mcp-execution-server → {mcp-execution-codegen, mcp-execution-introspector, mcp-execution-files, mcp-execution-core, mcp-execution-skill}
mcp-execution-codegen → {mcp-execution-introspector, mcp-execution-core}
mcp-execution-files → mcp-execution-codegen
mcp-execution-introspector → {rmcp, mcp-execution-core}
mcp-execution-skill → mcp-execution-core
See mcp-cli README for full reference.
# Generate TypeScript tools from config (recommended)
mcp-execution-cli generate --from-config github
# Preview what would be generated without writing to disk
mcp-execution-cli generate --from-config github --dry-run
# Introspect MCP server from config
mcp-execution-cli introspect --from-config github
# Introspect with detailed schemas
mcp-execution-cli introspect --from-config github --detailed
# Manual configuration (alternative)
mcp-execution-cli introspect docker --arg=run --arg=-i --env=TOKEN=xxx
# Shell completions
mcp-execution-cli completions bashTip
Use --from-config to load server settings from ~/.claude/mcp.json instead of manual arguments.
| Metric | Target | Achieved |
|---|---|---|
| 10 tools generation | <100ms | 0.19ms (526x faster) |
| 50 tools generation | <20ms | 0.97ms (20.6x faster) |
| VFS export | <10ms | 1.2ms (8.3x faster) |
| Memory (1000 tools) | <256MB | ~2MB |
- Progressive Loading Tutorial - Complete guide
- Claude Code Integration - Skill setup
- Architecture Decisions - ADRs explaining design choices
cargo build --workspace # Build
cargo nextest run --workspace # Test
cargo clippy --workspace # Lint
cargo +nightly fmt --workspace # FormatNote
All development follows Microsoft Rust Guidelines.
Minimum Supported Rust Version: 1.91
MSRV increases are considered minor version bumps.
Licensed under either of Apache License 2.0 or MIT license at your option.