Grafema

Grafema turns your codebase, infrastructure, knowledge, and workflows around it — into one queryable graph. For humans and AI.

We treat code as text. But text is just a form.

What actually matters when you write code is the system you have in your head — its structure. Entities, invariants, limitations. Goals and purpose. And how all these things relate to each other.

Software is naturally an executable graph — and so is everything around it: your services, your decisions, your team's knowledge. Grafema uses compiler-grade AST parsers — containing years of community-shared knowledge for each language — to excavate the deepest possible model of your system, and turn it into a transparent, queryable, enrichable map that grounds your understanding of it.

We refuse to accept "that's impossible to analyze statically." You can read code and understand it — you have a mental model in your head. So it's a matter of good enough heuristics. Human brains are literally built on this.

It's not magic and won't cover 100% of your system on day one. There will be gaps and "Here be dragons" signs. You will slay these dragons one by one — extend analysis with your own rules, fill up the knowledge base. And if you contribute, you slay one for everyone.

Thinking in graphs is not easy. But once it clicks - you stop reading code and just navigate the system. And your AI minions too.

Welcome to the party!

Licensed under FSL-1.1-Apache-2.0 — free to use, source available, converts to Apache 2.0 after 2 years. Details

v0.4.1 — Early access. Changelog | Known limitations

Quick Start

npm install -g grafema
grafema analyze --quickstart

That's it. --quickstart auto-detects your project languages, generates config, and builds the graph in one command.

For more control, use the two-step flow: grafema init (review config) → grafema analyze.

Explore your code

# What does this file do? (compact DSL overview, 10-20x smaller than source)
grafema tldr src/server.ts

# Who calls this function?
grafema who handleRequest

# Where does this data come from? (backward dataflow trace)
grafema wtf req.user

# Why is it structured this way? (knowledge base decisions)
grafema why auth-middleware

Use with AI (MCP)

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "grafema": {
      "command": "npx",
      "args": ["grafema-mcp", "--project", "."]
    }
  }
}

There is also a Docker image for running the MCP server (stdio) in a container — see the root Dockerfile: docker run -i --rm -v "$PWD":/workspace grafema-mcp.

30+ MCP tools available: find_nodes, find_calls, trace_dataflow, get_file_overview, describe, query_graph, and more. The AI agent queries the graph instead of reading files — faster, cheaper, more complete.

find_nodes returns rich context in a single call: callers, members, parent, import/call counts. Fuzzy name matching via local embeddings means approximate queries like find_nodes(name="PtyHostHeartbeatService") find HeartbeatService even without exact match.

Capabilities

Analyze

• ✅ Call graph — who calls what, across all files
• ✅ Data flow — trace values source to sink, forward and backward
• ✅ Control flow — CFG, reachability, branching paths
• ✅ Data shapes — object structure through assignment chains
• ✅ Effect propagation — transitive side-effect analysis through call graph
• ✅ Symbolic execution
• ✅ Cross-language & inter-process — service boundaries, message passing, remote calls
• ⏳ Side effect chain analysis
• ⏳ Inter-service contracts — message queue schemas, API schemas (OpenAPI, JSON Schema, gRPC)
• ⏳ Infrastructure as Code — Terraform, Kubernetes, Docker

Query

• ✅ CLI: tldr, who, wtf, why, check, overview
• ✅ 40+ MCP tools for AI agents (graph queries, navigation, dataflow, knowledge, git history)
• ✅ Datalog for custom structural queries
• ✅ Cypher query language
• ✅ Programmatic API (@grafema/util)
• ✅ HexAtlas — visual code map (2D/3D)
• ✅ VS Code extension

Document

• ✅ grafema export --as docs-md — generate human-readable docs from the live graph
• ✅ grafema export --as openapi-3.1 — auto-generate OpenAPI for HTTP routes
• ✅ grafema export --as mcp-schema — JSON-RPC tool registry, directly servable by any MCP runtime
• ✅ grafema export --as json-schema — Draft 2020-12 schemas per FEATURE
• ✅ Intent sidecars (_ai/intents/...) — handwritten "when to use" + captured examples that augment autogen output
• ✅ grafema features --duplicates — cross-modality dedup ("which CLI commands are wrappers around the same library function as which MCP tools")

Connect knowledge to code entities and flows

• ✅ Knowledge base — decisions, ADRs linked to code nodes
• ✅ Effects-DB & Registry — curated database of side effects and contract mappings for popular third-party packages across ecosystems (npm, PyPI, and more)
• ⏳ Git integration — blame, churn, authorship

Enforce your rules

• ✅ Architectural invariants as Datalog rules
• ✅ grafema check — CI gate
• ⏳ Code Quality Metrics — complexity, coupling, hotspots

Enrich with your own meaning

• ✅ Custom node types and edges via plugins
• ✅ Library callback enricher — auto-detect MCP tools, CLI commands
• ✅ Manifest generation — API surface with effect annotations

Language Support

JS/TS is the primary language with full dataflow support. Each language uses its community's canonical parser — not a generic tokenizer. grafema init includes all languages by default — analyzers for absent languages are simply skipped.

CLI Commands

VS Code Extension

Interactive graph navigation directly in your editor. Install from the VS Code Marketplace or search "Grafema Explore" in Extensions.

• Cmd+Shift+G — Find graph node at cursor
• Value Trace — See where data comes from and flows to
• Callers — All call sites for the function under cursor
• Blast Radius — Impact analysis: what breaks if you change this?
• Nodes in File — All graph nodes in current file with positions
• Explorer — Navigate edges (incoming/outgoing) interactively

Benchmarks

Analysis Performance

AI Agent Accuracy (Autoresearch)

Methodology: 30 questions sourced from real VS Code GitHub issues, scored by LLM judge. Questions span Sillito taxonomy levels L1 (finding focus) through L4 (full architecture understanding). Each question run as independent claude -p session with no prior context.