Skip to content

Super-Brain

Your codebase's working memory — local, fast, permanent.

Super-Brain gives your AI coding assistant a persistent, structured understanding of your code. Instead of re-reading files every turn, your assistant queries a local knowledge graph and gets back the exact functions, call paths, and semantics it needs.

Runs entirely on your machine. Plugs into 14 IDEs and AI coding tools. Zero code leaves your laptop.

The 30-second pitch

AI coding assistants have a context problem. Every session starts from zero. Every question re-reads the same files. Every long task burns tokens on exploration instead of answers.

Super-Brain fixes that. It ingests your code once, builds a call graph and vector index, and exposes it to your assistant via MCP. From then on:

  • "What calls processPayment?" → one graph query, millisecond response
  • "Explain the auth flow" → retrieve the actual path from HTTP handler to session store
  • "Where's this class used?" → transitive closure, not grep
  • "Summarize this module" → pre-clustered subsystems, already grouped

No cloud. No re-ingestion per session. No fabricated call relationships.

Why it's different

Capability Claude Code / Codex alone With Super-Brain
Call graph understanding Grep + read files Deterministic graph query
Persistent memory ❌ Session-scoped ✅ On disk, always available
Token cost per exploration Proportional to codebase Flat — queries, not reads
Privacy Uploads code to cloud Nothing leaves your machine
Cross-repo One project at a time Ingest multiple, query across
Binary/docs/audio Invisible First-class via ingest-doc / ingest-audio
Language coverage Top ~20 306 via tree-sitter

See the full comparison →

Install

pip install agsuperbrain

Or with uv:

uv add agsuperbrain

Full install guide →

Quick start

agsuperbrain init                    # configures + auto-ingests + indexes
agsuperbrain claude-install          # or: cursor-install, aider-install, etc.

One command sets up config, detects your source directory, builds the graph, embeds it for semantic search, and starts the background watcher. The second wires Super-Brain into your AI coding tool.

That's it. Your assistant now has permanent, structured access to your codebase.

Five-minute walkthrough →

Works with everything

Super-Brain ships install commands for 14 AI coding tools:

Claude Code · Cursor · Aider · Codex · OpenCode · GitHub Copilot CLI · VS Code Copilot Chat · Gemini CLI · Hermes · Kiro · Google Antigravity · OpenClaw · Factory Droid · Trae / Trae CN

One command per platform — installs the right hook, skill file, or rules into the right place. See the IDE matrix →

Any framework that speaks MCP also works: LangChain, LangGraph, AutoGen, CrewAI, SmolAgents, or plain stdio JSON-RPC.

What it gives you

  • 306 language support — Python, JS/TS, Go, Rust, Java, C/C++, Ruby, PHP, Kotlin, Swift, Scala, and 296 more via tree-sitter
  • Deterministic extraction — call graph comes from AST, not from an LLM guess
  • Hybrid retrieval — vector search for semantics + graph expansion for structure
  • Local LLM (optional) — answer questions using a bundled Llama-3.2-1B model, free and offline
  • Interactive visualization — Cytoscape.js graph, click to navigate
  • File watcher — code changes re-indexed automatically
  • Community detection — Leiden clustering surfaces subsystems without manual tagging
  • Multimodal — ingest PDFs, DOCX, MD, audio, and video into the same graph

Measured performance

From a self-corpus pilot (the agsuperbrain repository itself; reproducible via paper/evaluation/run_eval.py).

Token cost — same 10 questions, same LLM (Llama-3.3-70B):

Token cost: 210,659 vs 22,837

  • Without Super-Brain: 210,659 tokens total (mean 21,066 per question)
  • With Super-Brain: 22,837 tokens total (mean 2,284 per question)
  • 9.22× fewer tokens overall, comparable answer quality (within noise)

Extraction & retrieval:

  • 100% extraction per code-file on Python (40/40) and JavaScript (36/36)
  • 90–100% Section coverage on Markdown and HTML via the document pipeline
  • Sub-3 ms p95 on find_callers / find_callees graph queries; 7.4 ms p95 on transitive closure (depth 3)
  • 22.9 MiB on disk for the full code+doc index of a 5.4 MiB repository

Pilot scope, single-LLM caveat, and full numbers: see the paper.

Privacy by default

Every byte Super-Brain touches stays on your machine:

Nothing is uploaded. No telemetry. No account required.

Built by MEDHIRA

Super-Brain is part of MEDHIRA's mission to put engineering intelligence in the hands of developers, not vendors. About MEDHIRA →

Next steps