Conducks
CONDUCKSPROTOCOL

Structural code intelligence · No embeddings

Exact answers
about your code.

Conducks parses your repo into a structural graph with Tree-sitter, so AI agents and developers get graph-verified facts, not fuzzy embedding guesses.

Embedding search
“getUserById is somewhere in services…”
Conducks
getUserById · src/services/user.ts:42
called by 7 places · risk 0.31
Mirror, live architecture graph
View

Structural Intelligence

Structural
Intelligence

Conducks turns your source into a graph of symbols and the edges between them, calls, imports, inheritance, that you can query exactly.

Tree-sitter
DuckDB

Resilient extraction

When a grammar query fails to compile, Conducks falls back to a regex extractor, so a broken parse still yields file and symbol structure instead of nothing.

Weighted pathfinding

Finds the call and import paths between any two symbols using weighted edges, surfacing critical paths and blast radius across the dependency graph.

Graph-verified rename

Renames a symbol across every proven caller in the graph, with a dry-run preview before anything changes.

Detailed Mirror Analysis
View

Graph Detail

Graph
Detail

The detailed Mirror view exposes the edges of your codebase, the calls and imports that define how your system actually holds together.

Granular Mapping

Decompose complex inheritance chains and import graphs into actionable structural intelligence.

Sidebar Intelligence

Instant access to symbol gravity, risk scores, and blast radius metrics directly within the visual interface.

Workspace

Unified
Workspace

Conducks keeps the graph in sync with your editor as you work, so structural drift shows up the moment it happens.

Real-Time Sync

Changes in your editor are reflected in the graph right away, updating gravity and risk scores automatically.

Drift Detection

Conducks alerts you when a new import violates the structural rules you've set for your repository.

Engineering Workspace
View

Taxonomy

9-Layer
Taxonomy

Conducks maps every construct across all languages to a universal structural hierarchy, from ecosystems to atomic data points.

Universal Taxonomy Map
View
L0
Ecosystem
Cross-Repo
L1
Repository
Services
L2
Namespace
Modules
L3
Unit
Files
L4
Infra
Routes
L5
Structure
Classes
L6
Behavior
Methods
L7
Atom
Variables
L8
Data
Records

Language Support

Verified
Language Support

Four languages are fully extracted and verified, symbols, edges, frameworks, and entry points.

FeatureTypeScriptPythonRustGo
Imports
Named Bindings
Exports
Heritage
Type Annotations
Constructor Inference
Config
Frameworks
Entry Points

ExperimentalLenses present, extraction not yet verified, Java · C# · C · C++ · PHP · Ruby · Swift.

Governance

Agentic
Governance

The Conducks MCP server gives agents a 6-signal risk breakdown for any symbol, so your AI can weigh a change before making it.

Risk Decomposition

Gravity, Complexity, Entropy, Churn, FanOut, and Debt are distilled into a single actionable risk score.

Execution Tracing

Agents can trace execution paths to understand exactly how data flows through your system.

Agent Governance Hub
View

Architecture

Three-Layer
Architecture

Intelligence Layer

33 CLI commands and 13 MCP tools providing analysis, metrics, and governance.

Reflection Layer

Language-specific lenses powered by 14 Tree-sitter grammars and a regex fallback extractor.

Core Layer

Atomic DuckDB storage, weighted graph pathfinding, and the structural registry hub.

Intelligence
Reflection
Core Graph

Roadmap

What's Next

Our roadmap to the ultimate agentic governance platform.

Completed

Production TS / Python / Rust / Go
Real-time Mirror dashboard
13 MCP tools
Regex fallback extractor

In Progress

Query Template Library (19 templates)
Filter Builder for conducks_query
DAAC Community Clustering UI

Planned

import type Edge Detection (GVR gap)
Dynamic Dispatch Resolver
Deep Traversal Precompute

Get Started

Initialization

Get Conducks running in your local environment in a few minutes.

01

Build

$ git clone https://github.com/conducks/conducks
$ cd conducks
$ npm run bootstrap && npm run build
02

Connect

{
"mcpServers": {
"conducks": {
"command": "node",
"args": ["/absolute/path/to/conducks/build/src/interfaces/cli/index.js", "mcp"]
}
}
}

FAQ

Common
Questions

What is Conducks?

Conducks parses your source code with Tree-sitter into a structural graph stored in a local DuckDB vault. AI agents and developers can then ask exact questions about a codebase, where a symbol lives, what calls it, what breaks if it changes, and get graph-verified answers.

How is it different from embedding or RAG code search?

Embedding search returns approximate matches and can point to the wrong file or a symbol that does not exist. Conducks answers from a deterministic AST graph instead: exact symbol locations, real call and import paths, and no guessing.

Which languages does Conducks support?

TypeScript, JavaScript, Python, Rust, and Go are fully verified, symbols, edges, frameworks, and entry points. Java, C#, C, C++, PHP, Ruby, and Swift have experimental lenses where extraction is present but not yet verified.

Does my code leave my machine?

No. Conducks runs locally and stores the graph in a .conducks/ folder inside your project. There is no cloud, no account, and no embeddings sent anywhere.

How do AI agents use Conducks?

Through a Model Context Protocol (MCP) server that exposes 13 tools, query, trace, impact, explain, audit, rename, and more, so an agent can navigate and reason about your codebase over JSON-RPC.

Build
Better

GitHub Source