ctx graph

ctx graph builds and queries a symbol graph of your code. Answer questions that grep can’t answer cheaply.

Usage

# Index the project (once, then incremental)
ctx graph index --dirs src

# Who calls this function?
ctx graph callers run_proxy

# What does this function call?
ctx graph callees src/main.rs::execute

# What chain reaches this symbol?
ctx graph trace handle_request --depth 3

# What breaks if I change this?
ctx graph impact migrate_database

# Where is X defined?
ctx graph node CtxServer

Unique features vs CodeGraph

Feature	CodeGraph	ctx
Resolved call graph	✅	✅
Query-ranked results	❌ raw list	✅ BM25(query) + log(sites) + kind boost
Token budget on output	❌ can overflow context	✅ binary search like ctx map
Site deduplication	❌ N entries	✅ 1 entry + array of sites
Single Rust binary	❌ (TypeScript)	✅
Co-exists with hooks + MCP + RAG + repo map	❌ graph only	✅

Ranked example

ctx graph callers apply_pipeline --query "exec proxy" --max-tokens 800

Returns the top callers ranked by:

1. BM25 match between caller name/path and the query
2. Log of the number of call sites (popularity = relevance)
3. Boost for functions/methods over variables/constants

Languages supported (7)

Language	Extractor
TypeScript / TSX	✅
Python	✅
Ruby	✅
Go	✅
Rust	✅
Java	✅
Groovy	✅

Roadmap: C#, PHP, Swift, Kotlin, Scala, Dart, Vue, Svelte, Lua.

Storage

The graph is persisted in SQLite at ~/.cache/context_engine/graph.db. Schema:

symbols (id, name, qualified, kind, file, line, language)
calls   (id, caller_qualified, callee_name, file, line)
imports (id, file, module, alias)

Re-indexing a file clears its old entries first — fully idempotent.

Via MCP

Six MCP tools are exposed for agents:

• ctx_graph_index — populate the graph
• ctx_callers — who calls X?
• ctx_callees — what does X call?
• ctx_trace — caller chain up to depth
• ctx_impact — affected code on change
• ctx_node — definition location(s)