KNOWLEDGE LAYER FOR AI

YourAIfinally
understandsyour
codebase

Myceliums parses your entire codebase, extracts every symbol and relationship, and embeds it into a semantic knowledge graph. Your AI agent gets the full picture: every dependency, every call chain, every connection. It never misses context again.

54 MCP Tools·23 Languages·3 Search Modes·0 Cloud Dependencies
myceliumsNEW
Scroll

MCP Integration

Why do AI agents make 100+ blind grep calls when they could understand your entire codebase?

Give your AI agent real intelligence. Structured tools plus deep semantic context from the knowledge graph.

Myceliums analyzes your codebase, builds a semantic knowledge graph, and exposes it to Claude, Cursor, and Windsurf via MCP. Your agent doesn't just get tools, it gets understanding: every symbol, relationship, and dependency, ready to reason about.

ClaudeCursorWindsurfMCPmyceliums

Tool calls reduced

~80%

15+ grep calls → 3 KG queries *

Context size reduced

~99%

50K tokens → 500 tokens per query *

Complete results

100%

full call chain, no missed callers *

* Measured on standardized queries (caller lookup, blast radius, symbol search) across 6 open-source repos. See methodology

Basic Agent (100+ calls)OLD
User: 'Refactor getUserName safely'
Agent: Making blind grep calls...
$ grep -r 'getUserName' .
$ grep -r 'getUserName.*import'
$ grep -r 'function.*User'
$ grep -r 'class.*User'
... 96 more grep calls ...
Result: Still incomplete, lots of noise
Tokens spent: 50,000+ on context
KG Agent (5 calls)NEW
User: 'Refactor getUserName safely'
Agent: Using KG tools...
Tool 1: graph.getSymbol(name='getUserName')
Tool 2: graph.getCallers(symbol)
Tool 3: graph.getImpact(symbol)
Tool 4: graph.getRelated(symbol, depth=2)
Tool 5: code.getDefinition(symbol)
Result: Complete, accurate, safe refactor
Tokens spent: 2,000 precise context

Built for Agent Orchestration

What happens when you deploy not one agent, but ten, all working on the same codebase at once?

Every agent gets the same structured context. No conflicts, no blind spots, no duplicated work.

When an orchestrator dispatches multiple agents to a codebase, each one needs to understand the full architecture before making changes. Without a knowledge graph, every agent starts from scratch with grep and file reads. With Myceliums, they all share the same semantic understanding from the start: the complete call graph, community boundaries, and dependency map. The orchestrator assigns, the agents execute with precision.

Knowledge Graph
Refactor Agent
Test Agent
Review Agent
Docs Agent

Shared context

1 graph

All agents query the same knowledge graph. No redundant indexing.

Conflict prevention

0 overlaps

Impact detection shows each agent exactly what other agents might affect.

Parallel execution

N agents

Refactor, test, review, and document simultaneously on independent community boundaries.

Architecture

What happens when you run myc analyze?

From source files to AI-ready knowledge in 8 stages. All local, all Rust, all under a second for queries.

Your code passes through a pipeline of parsing, graph construction, community detection, and indexing. The result is a structured knowledge graph that any AI agent can query via MCP or CLI.

Source Files23 languagesYour codebase
tree-sitterParserAST extraction
Symbol StoreLanceDBVectors + metadata
Graph BuildRelationshipsCalls, imports, members
CommunitiesLeidenArchitecture clusters
Search IndexBM25 + VectorsHybrid retrieval
Process TraceEntry→ExitExecution flows
AI AgentsMCP / CLI54 tools exposed
Rust Workspace6 crates
myc
CLI binary
myceliums-mcp
MCP server
myceliums-core
Parsing, analysis, search
myceliums-storage
LanceDB persistence
myceliums-cypher
Cypher query engine
myceliums-http
Visualization server

All local. No cloud dependencies. Zero telemetry. Written in Rust for maximum performance and safety.

// ingest.surface

20+ languages and every format you work with.

AST-level extraction across the full polyglot stack. Docs, notebooks, PDFs, and emails. One corpus, one graph.

PythonTypeScriptJavaScriptGoRustJavaC/C++C#RubySwift
KotlinPHPLuaZigScalaElixirObjective-CPowerShell
PDFMarkdownMDXJupyterEmailPlain Text
PythonJS/TSGoRustJavaC/C++RubySwiftKotlinPHPC#ScalaElixirPDFMDJupyterEmail

What Myceliums can do

15 capabilities built into a single binary. Click any card to see it in detail.

From parsing and graph construction to impact detection, search, and AI agent tooling.

Knowledge Graph

Parses 23 languages into a structured graph of symbols, relationships, and call chains.

Impact Detection

Analyze the blast radius of any change before shipping. Traces through calls and imports to find every affected function and process.

Structure, not guesswork

Organized relationships between functions, types, and modules. Ready to reason about, not raw file contents.

export function handleAuth(req) {
const token = validate(req);
if (!token) throw new Error();
return session.create(token);
}
import { TokenStore } from './store';
handleAuthvalidatesessionTokenStorecreate

Community Detection

Leiden algorithm groups related code into logical clusters. Modularity scoring, cohesion metrics, and inter-community coupling reveal architectural quality.

Process Tracing

Maps execution flows from entry points through entire call chains automatically.

mainauthapivalidatesessionhandlerdb

Cypher Queries

Query your codebase like a database with the same language as Neo4j.

|

MCP Integration

54 tools for AI agents, with one-command setup for 11 editors and agents. Claude Code, Cursor, Copilot, Gemini, and more via Model Context Protocol.

ClaudeCursorWindsurfMCPmyceliums

Hybrid Search

BM25 keyword + semantic vector search with Reciprocal Rank Fusion for the best of both worlds.

$|

Token-Optimized

5-20x fewer tokens for code review. Structural summaries instead of full files.

Before142,000 tokens
After2,800 tokens
50x reduction

Interactive Viz

Force-directed, hierarchical, and radial graph layouts with community drill-down, metrics overlay, and SVG export.

myc serve

Local-First

No cloud. No telemetry. No API calls. Your code never leaves your machine.

No CloudNo Telemetry

Circular Dependency Detection

Detect mutual imports and call cycles via Tarjan's SCC algorithm. No other tool does this automatically.

Cycle 1 (3 symbols, 2 files):
auth → billing → auth
Cycle 2 (2 symbols, 1 file):
parse → validate → parse

Centrality Metrics

Four centrality metrics — betweenness, closeness, eigenvector, degree — identify bridge symbols, bottlenecks, and critical paths.

# Name Between Close
1 processReq 0.8421 0.7103
2 authMiddle 0.6234 0.6891
3 dbConnect 0.4102 0.5540

Module Coupling

Afferent/efferent coupling and instability metrics per module. Find fragile code and transitive dependency chains.

Module Ca Ce Instability
src/api 2 8 0.8000
src/auth 5 3 0.3750
src/core 9 1 0.1000

Design Rationale

Extracts WHY/TODO/HACK comments and links them to the code they explain. Know the reasoning, not just the structure.

// HACK: temporary workaround for #142
→ linked to: processPayment()
// TODO: add retry logic here
→ linked to: fetchUserData()

Architectural Intelligence

God nodes, surprising connections, and auto-generated reports surface hidden bottlenecks and coupling.

God node: AppRouter (47 connections)
Surprising: auth → billing (0.92)
Report: 3 bottlenecks, 2 isolated

Knowledge Gaps

Finds untested code, isolated modules, and documentation holes automatically.

Untested: handlePayment() (8 callers)
Isolated: legacy/ (1 external edge)
No docs: src/core/ (23 symbols)

Multi-Format Export

Export to GraphML, Neo4j Cypher, SVG, JSON, or generate an Obsidian wiki from your knowledge graph.

myc export --format graphml
myc export --format cypher
myc export --format svg
myc wiki --output docs/

Always up to date

Git hooks auto-update on every commit. Graph diff tracks what changed. Your understanding stays current.

main.tsauthapistoreroutes

Graph Diff

Compare graph snapshots to track how your architecture evolves over time.

+12 symbols added
-3 symbols removed
+8 relationships
2 communities changed

// architecture intelligence

Your architecture, always visible

Once the knowledge graph is built, Myceliums becomes your architecture guardian — visualizing services, detecting drift, and flagging quality issues automatically.

Architecture Lint

4 rules: cycles, god nodes, fan-out, instability

NEW

Quality Hotspots

Centrality x churn x instability = refactor priority

NEW

All features available as MCP tools — works with Claude Code, Cursor, Windsurf, and any MCP client.

10 architecture intelligence capabilities — all available as MCP tools

NEW

Architecture Diagrams

Service-level views from communities

NEW

Mermaid Export

Flowchart, class, and graph diagrams

NEW

Quality Hotspots

Centrality x churn x instability

NEW

Architecture Lint

Cycles, god nodes, fan-out checks

NEW

Drift Detection

Track architectural changes over time

NEW

Snapshot Versioning

Timestamped graph history + diff

NEW

CODEOWNERS

Auto-resolve file ownership

NEW

Service Mapping

Name your communities as services

NEW

Decision Records

ADRs linked to code symbols

NEW

API Contracts

OpenAPI + Proto handler matching

Knowledge Graph

How can you understand any codebase without manually reading every file and tracing relationships by hand?

Understand any codebase in seconds. 20+ languages parsed into a complete map of every symbol, relationship, and call chain.

Every symbol, relationship, and call chain across your entire codebase — structured, queryable, and ready for AI agents.

Languages

20+

from manual parsing to automatic

Symbols parsed

38K

in Django alone

Query time

0.003s

vs hours of manual tracing

Live Preview

INTERACTIVE
Initializing force simulation...

Force-Directed Layout

D3.js simulation with charge, link, collision, and clustering forces. Nodes naturally organize by their relationships.

Community Coloring

Leiden algorithm clusters rendered as colored groups. See how your code actually connects vs. what folders say.

Centrality Scaling

Node size reflects betweenness centrality. Larger nodes = more critical. Spot bottlenecks instantly.

Cycle Highlighting

Circular dependencies rendered in red with dashed edges and warning rings. Tarjan's SCC algorithm detects them.

Without Knowledge GraphOLD
$ grep -r "className=\"button\"" --include="*.tsx"
src/ui/Button.tsx: export const Button = ()
src/Button.tsx: export function Button()
src/components/Button.tsx: const Button = ()
$ find . -name '*.tsx' | xargs grep -l 'Button'
src/pages/Login.tsx
src/pages/Signup.tsx
src/pages/Dashboard.tsx
src/pages/Settings.tsx
// Still missing which functions use which Button?
// How are they imported? What's the call flow?
// Manual investigation required... Hours of work
With Knowledge GraphNEW
$ myc query "MATCH (s:Symbol)
WHERE s.name = 'Button'
RETURN s.name, s.file, count(*) as callers"
┌──────────┬──────────────────┬─────────┐
│ s.name │ s.file │ callers │
├──────────┼──────────────────┼─────────┤
│ Button │ src/ui/Button │ 47 │
└──────────┴──────────────────┴─────────┘
✓ 1 query. Complete graph. 0.003s

Run myc serve in your project to see this for your own codebase

Incremental Intelligence

Why do RAG pipelines re-embed everything on every change?

Build once, grow forever. Only affected nodes update. Your knowledge graph is always current, never stale.

Myceliums maintains a living graph. When a file changes, only the affected subgraph is patched. No full rebuilds, no stale context. AI agents and developers always work with the real state of your code.

// day.one

Initial corpus build

Point Myceliums at any directory. AST extraction, search indexing, community detection. Your entire codebase becomes a traversable knowledge graph in one pass.

5 nodes · 6 edges

// week.one

Incremental updates

A file changes. Myceliums detects it, re-extracts only the affected subgraph, and patches the live graph. No full rebuild. Your knowledge stays connected.

+6 nodes patched in · 0.8s

// month.one

Enterprise-scale corpus

After weeks of continuous growth, your graph spans thousands of symbols across dozens of modules. Query traversal is instant. Nothing was ever rebuilt from scratch.

3 communities · 21 nodes · enterprise corpus

// myc watch . — live file watcherno LLM · AST only

// others: full rebuild on change

re-embed 500k documents...4h 12m
rebuild vector index...2h 8m
re-cluster communities...stale context

// myceliums: incremental patch

detect changed files...3 files
patch affected nodes only...0.8s
graph updated. 498,752 nodes intact.
RAG / grep approachGeneric retrieval, blind to structure
blind-searchOLD
$ grep -r 'handleAuth' --include='*.ts'
src/auth/handler.ts: export function handleAuth(req) {
src/auth/handler.ts: const token = validateToken(req);
src/middleware/auth.ts: import { handleAuth } from '../auth';
 
$ grep -r 'validateToken' --include='*.ts'
src/auth/validator.ts: export function validateToken(token) {
src/auth/handler.ts: const token = validateToken(req);
 
$ grep -r 'TokenStore' --include='*.ts'
src/storage/tokens.ts: class TokenStore {
src/auth/validator.ts: import { TokenStore } from '../storage';
 
... 47 tool calls later ...
// still missing 3 callers
MyceliumsStructural queries, complete results
myceliumsNEW
$ myc query "MATCH (s)-[:CALLS*1..3]->(t)
WHERE t.name = 'handleAuth'
RETURN s.name, s.file"
 
┌─────────────────────┬──────────────────────┐
│ s.name │ s.file │
├─────────────────────┼──────────────────────┤
│ authMiddleware │ src/middleware/auth │
│ loginRoute │ src/routes/login │
│ oauthCallback │ src/routes/oauth │
│ sessionRefresh │ src/auth/session │
│ apiGateway │ src/gateway/main │
└─────────────────────┴──────────────────────┘
 
✓ 1 query. Complete call chain. 0.003s
USE CASES

Built for real workflows

Pre-deploy Safety

Run impact detection on every PR. Know exactly which functions, communities, and processes are affected before merging.

$myc impact --diff HEAD~1

Codebase Onboarding

New team members understand architecture in minutes. Process tracing shows how data flows, communities show boundaries.

$myc processes ./my-project

Architecture Audits

Community detection reveals actual module boundaries. Find hidden coupling, circular dependencies, and god classes.

$myc communities ./my-project

AI-Powered Code Review

Token-optimized context gives AI agents structured understanding instead of raw files. 50x fewer tokens, better reviews.

$myc review --diff HEAD~1

Documentation Generation

Generate an Obsidian vault from your knowledge graph. Every community becomes a page with symbols, relationships, and cross-references.

$myc wiki ./my-project --output docs/

AI Review Intelligence

Auto-generated review questions highlight risky changes, god nodes, and missing tests before you merge.

$myc report ./my-project
100% LOCAL · ZERO TELEMETRY · Apache-2.0 License

Privacy by architecture, not by promise

Local-First Architecture

  • All processing on your machine
  • LanceDB embedded storage
  • No network calls during analysis

Open Source & Auditable

  • Apache-2.0 licensed
  • Full source on GitHub
  • Every claim here is verifiable in the code

Zero Telemetry

  • No usage tracking
  • No analytics
  • No phone-home behavior

No Vendor Lock-In

  • Standard formats
  • Exportable graph data
  • Works offline forever

You choose where your data goes

🏠

Local AI

Ollama, HuggingFace, local models

Your code → Your machine
→ Your local LLM
→ Answers stay internal
Nothing leaves your network
☁️

Cloud AI

Claude, ChatGPT, Gemini APIs

Your code → Your machine
→ API call (encrypted)
→ Query data leaves network
You control what gets sent
🔒

No AI needed

Pure structural matching

Your code → Your machine
→ Regex + graph traversal
→ Zero external calls
Works fully offline
Apache-2.0 LicenseAudit the code yourself →

How Myceliums compares

Code knowledge graphs got crowded fast, and several of the alternatives below are far more established than this one. Here is where Myceliums actually differs — and where you should pick something else.

FeatureMyceliumsUAGLabCRGTSGraphifyGreptileCodeRabbit
Cypher query engine (read-only)
Process / execution-flow tracing
Impact detection before a change
Symbol rename preview
Architecture drift detection
BM25 + semantic + hybrid search
MCP server
Local-first (no cloud)
Indexes without an LLM
Community / god-node detection
Ingests docs, configs, SQL schemas
Automated PR review in CI
Persistent cross-session memory
Proven at scale / large user base
Commercial support / SLA
PriceFreeFreeFree*FreeFreeFree*$30/seat*$24/user
ImplementationRust CLITypeScriptRustPythonPythonPythonCloudCloud

Pick something else if: you want the most-adopted option with the largest community — Graphify and Understand-Anything are both far ahead there. If you need code review posted into your PRs with a support contract behind it, Greptile or CodeRabbit are products; this is a tool. And if your pain is scattered documentation rather than code structure, Graphify handles non-code sources better today.

Compiled from public documentation and repositories, July 2026. Competitors ship quickly and this will go stale — if a row is wrong, please open an issue and it gets corrected. * GitLab KG requires a GitLab instance (free tier available). Graphify's enterprise tier is early access with no public pricing. Greptile is $30/seat plus $1 per review beyond 50.

Benchmarks

Does it actually work on real codebases at scale?

Tested on 6 popular open-source repositories across 6 languages. The raw data is in the repo — check it.

0%Avg Token Reduction
0%Avg Call Reduction
0Repos Tested
0Queries Validated
django
Python82k
6,492
Files
38,210
Symbols
52,417
Rels
~90m
First run
35.7s
Updates
Token reduction99.8%
Call reduction94.8%
fastapi
Python82k
1,118
Files
5,330
Symbols
8,089
Rels
9m
First run
0.9s
Updates
Token reduction99.7%
Call reduction92.8%
zod
TypeScript35k
361
Files
7,327
Symbols
8,403
Rels
11m
First run
0.8s
Updates
Token reduction99.9%
Call reduction94.7%
express
JavaScript66k
498
Files
2,841
Symbols
3,672
Rels
4.5m
First run
0.2s
Updates
Token reduction99.8%
Call reduction93.4%
gin
Go80k
312
Files
1,847
Symbols
2,518
Rels
~3m
First run
0.1s
Updates
Token reduction99.6%
Call reduction93.7%
tokio
Rust28k
1,024
Files
8,493
Symbols
11,247
Rels
24m
First run
0.3s
Updates
Token reduction99.7%
Call reduction94.7%
Methodology & raw data on GitHub →

File, symbol and relationship counts and the token and call reductions come from the published benchmark data linked above. The run times are indicative estimates from development runs on an Apple M1 Pro — they are not part of that dataset and will vary with hardware, embedding model, and repository. First run includes embedding generation; updates only re-analyze changed files.

Get started now

Ready to give your AI agents real understanding of your codebase?

Two commands. Myceliums installs, auto-detects your editors, and configures everything.

1

Install

One command. Works on Mac, Linux, and Windows. Also available via Homebrew and Docker.

$ cargo install myc
2

Setup

Auto-detects all installed editors and configures MCP integration for each one.

$ myc setup

What happens under the hood

1myc setupauto-detect + configure
$ myc setup
Detecting installed editors...
✓ Claude Code configured (MCP + hooks)
✓ Cursor configured (MCP)
✓ VS Code configured (MCP)
✓ Windsurf configured (MCP)
✓ Zed configured (context server)
- JetBrains not found
- Aider not found
5 editors configured. Run myc doctor to verify.
2myc analyze .build knowledge graph
$ myc analyze .
Parsing 5,247 files across 20+ languages...
Building knowledge graph...
Symbols: 41,952
Relationships: 45,028
Communities: 86
Processes: 214
✓ Analyzed in 1.6s
$ myc search "handleAuth"
1. handleAuth (Function) — auth.ts:14 [0.98]
2. authMiddleware (Function) — middleware.ts:31 [0.87]

Supported editors and AI agents

Claude CodeCursorVS CodeWindsurfZedJetBrainsCopilot CLIGemini CLICodex CLIAiderKiroContinue

myc setup auto-detects which editors are installed and configures each one. Claude Code gets extra hooks for auto-indexing on every commit.