One binary. 13 tools. Built to make AI agents faster, cheaper, and more precise.
cortex-works-minimal is a lean MCP server for AI-powered IDEs. It replaces slow, token-hungry IDE tooling with a disciplined surface of 13 surgical tools — covering repo mapping, symbol analysis, structural editing, filesystem work, and bounded shell execution.
A hyper-optimized MCP server that supercharges AI coding agents. Outperforms built-in IDE tools with AST-based precision editing and native OS control, drastically reducing LLM token consumption.
星數
8
工具
—
可用率
—
✅ Cross-platform tested: macOS, Windows, and Ubuntu (other Linux distros may work too).
Setup
Before building on macOS, make sure the Rust toolchain, Apple command line tools, and CMake are installed:
If you already have the Apple command line tools, you can skip the install step. This repository now uses the default macOS linker, so you do not need a Homebrew-specific lld path to build the release binary.
cargo build --release -p cortex-mcp
./target/release/cortex-mcp # wire to your IDE via MCP stdio
Why Agents Perform Better Here
Standard IDE tools expose the filesystem as a flat text surface. Agents waste tokens reading full files to find the three lines they need, burning context on boilerplate and build artifacts that should never appear in a prompt.
cortex-works is designed around the opposite principle: return exactly what the agent needs, nothing more.
Token Efficiency by Design
The stack enforces a layered, progressive-disclosure model:
Stage
Tool
What the agent learns
L1 — Discover
workspace_topology
Roots, manifests, language hints — <500 tokens
L2 — Map
map_overview / skeleton
File tree + symbol list for selected dirs only
L3 — Read
read_source / deep_slice
Exact symbol body or focused file slice
An agent that follows this pattern typically spends 5–10× fewer tokens on codebase orientation than one using raw file reads.
Structure-Aware Edits, Not Line Patches
Line-number-based edits break as soon as another tool touches the same file. cortex-works edits by name:
cortex_act_edit_ast — replace a Rust/TS/Python function or struct by symbol name
cortex_act_edit_data_graph — update a JSON or YAML key by JSONPath target
cortex_act_edit_markup — rewrite a Markdown section by heading or HTML node by tag/id
cortex_act_sql_surgery — swap a DDL statement by type and object name
The agent supplies the new content; the tool handles byte offsets, bottom-up ordering, and post-edit validation automatically.
Fail-Safe AST Editing
cortex_act_edit_ast validates syntax after edits and aborts safely if the output is invalid. This avoids writing broken code to disk.
Multi-Root Path Routing, Built-In
VS Code multi-root workspaces, Zed multi-project, and JetBrains polyrepo sessions all work transparently via prefixed path convention: [FolderName]/path/to/file. No manual path juggling required.
One Round-Trip for Sequential Workflows
cortex_act_batch_execute collapses multi-step workflows — explore → checkpoint → edit → verify — into a single MCP call, cutting back-and-forth latency by up to an order of magnitude.
The 13 Active Tools
Intelligence (read-only)
Tool
Use when
cortex_code_explorer
First look at an unfamiliar repo; topology, maps, deep slices
cortex_symbol_analyzer
You already know the symbol; read source, find usages, blast radius
cortex_chronos
Save a rollback point before risky refactors; compare before/after
cortex_manage_ast_languages
A non-core parser is missing (Go, PHP, Ruby, Java, C, C++, C#, Dart)
Structural Mutations
Tool
Use when
cortex_act_edit_ast
Editing a Rust, TypeScript, or Python symbol by name
cortex_act_edit_data_graph
Updating JSON or YAML keys structurally
cortex_act_edit_markup
Rewriting Markdown sections, HTML/XML nodes by heading/tag/id
cortex_act_sql_surgery
Replacing a DDL statement (CREATE TABLE, CREATE INDEX, …)
cortex_fs_manage
Creating, copying, moving, deleting files and dirs; patching .env/.ini
Search, Execution, and Runtime
Tool
Use when
cortex_search_exact
A literal string, regex, or identifier you know exactly
cortex_act_shell_exec
Short diagnostic commands; manifest-aware run_diagnostics mode
cortex_act_batch_execute
Collapsing a sequential workflow (explore → edit → verify) into one call
cortex_mcp_hot_reload
After rebuilding, reload the worker on the same stdio channel without restarting the IDE
Quick Start
# 1. Clone and build
git clone https://github.com/cortex-works/cortex-works-minimal
cd cortex-works-minimal
cargo build --release -p cortex-mcp
# 2. Wire to VS Code (add to mcp.json)
{
"cortex-works": {
"type": "stdio",
"command": "/absolute/path/to/target/release/cortex-mcp"
}
}
VS Code Extension Wrapper
The repository now also includes a VS Code workspace extension wrapper in extensions/cortex-works-vscode.
It now does one job cleanly: expose the same 13 Cortex tools through an extension-native bridge instead of a user-configured MCP server entry.
no extension-only tool names are added
the user-facing tool surface stays aligned with cortex-mcp
the extension runs cortex-extension-bridge, which reuses the existing Cortex dispatch directly for lower overhead than the MCP path
A hyper-optimized MCP server that supercharges AI coding agents. Outperforms built-in IDE tools with AST-based precision editing and native OS control, drastically reducing LLM token consumption.
上次檢查時間:2026-10-28。 A hyper-optimized MCP server that supercharges AI coding agents. Outperforms built-in IDE tools with AST-based precision editing and native OS control, drastically reducing LLM token consumption.