Mycelium — Codebase memory for AI agents
AI codebase intelligence
Your agent reads<br>40 files. It should<br>read 4.
One command gives your AI agent persistent, queryable memory of your entire codebase — before it touches a single file.
npx @kopikocappu/mycelium init<br>copy
Works with Claude Code · Cursor · GitHub Copilot · Any MCP-compatible agent
↓ scroll
10×
Fewer files read
7k
Tokens saved per task
70+
Files mapped per project
Command to start
The problem
AI agents are powerful. But they're flying blind.
Without Mycelium
Your agent guesses.
Before writing a single line, your agent opens file after file trying to build a mental model it has never had. By the time it starts, it's read 40 files and lost context on most of them.
app/page.tsx
app/layout.tsx
lib/auth-context.tsx
lib/user.ts
components/Navbar.tsx
components/ProGate.tsx
app/api/stripe/webhook/route.ts
+ 33 more files...
With Mycelium
Your agent knows.
Before touching anything, your agent calls /preflight and gets back exactly the files it needs — with plain English descriptions of what each one does.
app/api/checkout/route.ts
app/api/stripe/webhook/route.ts
lib/user.ts
app/pro/page.tsx
How it works
Three steps. Then your agents are smarter forever.
01 — SCAN
$ mycelium init
Map everything
Mycelium parses every file, resolves every import, and builds a dependency graph of the entire codebase. Claude Haiku writes a plain English description of what each file does and why it exists.
02 — QUERY
GET /preflight?task=
Ask before touching
Your agent calls /preflight with a plain English description of its task. Mycelium returns exactly which files matter and how they connect. Four files, not forty.
03 — TRACK
GET /history
Know what happened
Every file save is logged with a timestamp, a task description, and which agent made the change. You always know what your agents did — and when.
Features
Everything your codebase never told your agents.
Natural language preflight
Describe your task in plain English. Get back exactly the files that matter, with AI-written explanations of what each one does and how they connect.
Live graph viewer
A browser-based visualization of your entire codebase. Files as nodes, imports as edges, call relationships in orange. Understand any codebase in five minutes.
Agent change history
Every file save logged by task and agent identity — human, Claude Code, Cursor, anything. The full audit trail of what your AI agents actually did.
Semantic search
Find files by concept, not keyword. "Authentication flow" finds the right files even when none of them have the word "auth" in their name.
Auto-bootstrap
Mycelium writes CLAUDE.md and .mcp.json into your project automatically. Agents discover it and use it without any manual configuration.
Cross-reference analysis
Before modifying a function, ask what calls it. Get the full blast radius — every file and function that depends on what you're about to change.
In practice
What your agent sees before it opens a file.
An agent about to implement a Stripe checkout calls /preflight and gets back exactly the right context. No guessing. No wasted reads. Just the graph knowing what matters.
The server runs locally at localhost:47821 — your code never leaves your machine.
mycelium — preflight
curl "localhost:47821/preflight?task=add+stripe+checkout"
"task": "add stripe checkout",
"mode": "semantic",
"tokensSaved": 7000,
"files": [
{ "nodeId": "app/api/checkout/route.ts", "score": 0.94 },
{ "nodeId": "app/api/stripe/webhook/route.ts", "score": 0.91 },
{ "nodeId": "lib/user.ts", "score": 0.87 },
{ "nodeId": "app/pro/page.tsx", "score": 0.82 }
// 4 files. 7,000 tokens saved. Agent starts writing.
Who it's for
Built for everyone who ships code with AI.
Developers
Make your agent actually understand what it's doing.
Stop watching Claude Code open 30 files before writing a single line. Mycelium gives it a map. It reads 4 files and gets it right.
Claude Code users
Cursor users
GitHub Copilot users
Any MCP-compatible setup
Engineering leads
Finally know what your agents are doing to the codebase.
Every agent session logged. Every file touched, timestamped, attributed. The audit trail that AI coding tools forgot to build.
Session-level history
Per-agent attribution
Task-based grouping
Full change timeline
New engineers
Understand any codebase in five minutes.
Open the graph viewer. See every file, every dependency, every connection — with plain English explanations on every node. No more codebase spelunking.
Visual dependency map
AI descriptions on every file
Semantic search by concept
Works on any language
Get started
Your codebase has been waiting to be understood.
One command. Any project. Your agent never files blind again.
npx @kopikocappu/mycelium init<br>copy
View on GitHub<br>npm package
🍄 mycelium
Built by Minh Tran · MIT License · 2026