Field Guide: A Taxonomy of Code-Context Tools

The problem

Why a feature grid lies

A row like "can it write files?" scores a read-only retriever as an empty cell — the toaster losing at air-frying. But read-only isn't a missing feature; for an agent it's a deliberate smaller attack surface. The grid flattens a design choice into a deficit.

The fix: declare the appliance before you compare.

Nobody's confused when you say "a toaster and an air fryer both apply dry heat to food; they differ at heat source, chamber geometry, airflow, and target food." Name the axes and the comparison becomes honest. Every tool below is the same six-stage pipeline making different choices — so we name the class, then profile the stages.

Axis 1

The operational pipeline

Every code-context tool is this pipeline. A tool's identity is just its column of choices down it.

Stage 1

Ingestion

How is source acquired & parsed?

Stage 2

Representation

What artifact gets stored?

Stage 3

Retrieval

How do you ask?

Stage 4

Grounding

Is output tied to real source lines?

Stage 5

Action

What can it do to the code?

Stage 6

Delivery

How does it reach the agent?

Axis 2

Componentry — the parts list

Orthogonal to the stages: the physical parts a tool is built from. Two tools can share a stage choice yet differ sharply in parts (a graph in SQLite vs a graph in KuzuDB).

Parser tree-sitter · LSP · native compiler Store SQLite index · vector DB · graph DB · flat file Ranker heuristic · embedding · learned weights Runtime none · write · execute-sandbox Scope single-repo · suite · org-rollup Transport stdio · SSE · HTTP

The appliance classes

Appliance classes, declared up front

Before any scoring, every direct alternative sorts into one of these. The class tells you what the tool is for — and what it deliberately trades away to be good at it.

Grounded read-only retriever

For: Answer "where / what is this symbol" with line-exact source, spending the fewest tokens possible.

Trades: Deliberately cannot write or run code. Not a semantic-similarity search — it returns what is there, not what is like it.

jCodeMunch cymbal trace-mcp vexp SigMap Pharaoh

Graph-database code intelligence

For: Model the codebase as a graph (call edges, imports, entities) and traverse relationships.

Trades: A graph build and store to maintain; answers are as good as the graph, and edges can drift from source.

GitNexus Dual-Graph (GrapeRoot) code-review-graph Context+ Axon Octocode CodeGraph codebase-memory-mcp

Embedding / RAG search

For: Find code by meaning — "the thing that handles auth" — via vector similarity.

Trades: Recall over fidelity: results are ranked probabilities, not guaranteed line-exact facts. Needs an embedding step and a vector store.

SocratiCode

Whole-repo packer

For: Flatten the whole repo (or a budgeted map of it) into one artifact and hand it to the model in a single shot.

Trades: No query-time retrieval; the model pays to read the pack. Great for small repos, expensive as they grow.

Repomix RepoMapper

Mutate / execute runtime

For: Give the agent hands: read AND write files, run tools, refactor, apply edits.

Trades: A different appliance entirely. Larger attack surface and less predictable behaviour; retrieval is one feature among many.

Raw File Tools mcp-server-filesystem Serena SDL-MCP TokenSave

Context interception layer

For: Sit between the agent and the model, compressing what gets sent — prompt, history, tool results, shell output — so the bill falls without the agent changing.

Trades: The largest trust surface of any class here: it reads and rewrites every request in flight. Savings are real, but so is the failure mode — a compressor that drops a line drops it silently, and the agent cannot know what it did not receive.

LeanCTX

The stage fingerprint

Every tool, profiled down the pipeline

Grouped by class, then profiled across all six stages. Where an axis is honestly graded, colour shows it; Action is a class declaration, not a score (see the caveat below).

line-exact facts tied to real lines probabilistic ranked similarity synthesized generated summary/map read-only write / execute

Tool	IngestionHow is source acquired & parsed?	RepresentationWhat artifact gets stored?	RetrievalHow do you ask?	GroundingIs output tied to real source lines?	ActionWhat can it do to the code?	DeliveryHow does it reach the agent?
GROUNDED READ-ONLY RETRIEVER
jCodeMunch	tree-sitter AST	SQLite symbol index	symbol + ranked hybrid	line-exact	read-only	consolidated, token-annotated
cymbal	tree-sitter	SQLite / FTS5	symbol lookup	line-exact	read-only	Go CLI
trace-mcp	tree-sitter (AST-first)	symbol index	symbol + flow trace	line-exact	read-only	~50-tool default since 2026-07-26 (~170 on opt-in full)
vexp	Rust parser	local index	hybrid context	line-exact	read-only	native binary
SigMap	parse	ranked file list + signature file	query → ranked files	synthesized signatures	read-only (writes artifact)	file artifact, not agent tools
Pharaoh	open AST parser	hosted layer	symbol	line-exact	read-only	OSS parser + hosted service
GRAPH-DATABASE CODE INTELLIGENCE
GitNexus	graph build	knowledge graph	graph traversal	graph-derived	read-only	"nervous system" framing
Dual-Graph (GrapeRoot)	parse	dual graph	graph traversal	graph-derived	read-only	very active release cadence
code-review-graph	tree-sitter	persistent SQLite graph	graph traversal	line-linked	read-only	review-oriented
Context+	TS parse	hierarchical feature graph	graph + search	synthesized	read-only	feature-graph framing
Axon	parse	KuzuDB graph + vector	hybrid (BM25 + vector + fuzzy)	probabilistic	read-only	community detection (Leiden)
Octocode	tree-sitter	LanceDB vector + typed graph	hybrid	probabilistic	read-only	Rust binary + CLI
CodeGraph	TS parse	pre-indexed knowledge graph	graph lookup	line-linked	read-only	ships prebuilt to Claude Code
codebase-memory-mcp	parse	persistent knowledge graph	graph	probabilistic	read-only	claims learning / memory
EMBEDDING / RAG SEARCH
SocratiCode	AST-aware chunking (ast-grep)	per-branch Qdrant vectors	dense vector + BM25	probabilistic	read-only	enterprise vector intel
WHOLE-REPO PACKER
Repomix	read files	flat file (XML / MD / text)	one-shot dump	verbatim (whole repo)	read-only	single artifact
RepoMapper	tree-sitter	token-budgeted map	map generation	synthesized map	read-only	single map artifact
MUTATE / EXECUTE RUNTIME
Raw File Tools	none (raw)	none	grep / glob	line-exact (whole file)	read + write (bash)	native to the agent
mcp-server-filesystem	none	none	path / list / search	line-exact	read + write	13 filesystem ops
Serena	LSP	in-memory (LSP)	semantic symbol (IDE-level)	line-exact	read + write + execute	coding-agent toolkit
SDL-MCP	tree-sitter (Rust indexer)	embedded graph-DB	graph + symbol delta	line-exact	read + write + execute	gateway + delta ledger
TokenSave	tree-sitter (50+ langs)	libSQL symbol index	symbol + search	line-exact	read + write	~80 tools, token TUI
CONTEXT INTERCEPTION LAYER
LeanCTX	tree-sitter AST (27 langs)	property graph + embeddings	BM25 + vector, RRF with graph proximity	line-exact on read, rewritten in transit	read + intercept + compress	MCP tools plus an optional provider proxy

jCodeMunch's row is the only one that is line-exact grounding + read-only + a consolidated, token-annotated surface at the same time.

Intellectual honesty

Where the taxonomy stops being a ranking

Two of the six stages are incommensurable — you can describe them, but you can't honestly rank them. Pretending otherwise is exactly the air-fryer/toaster mistake in reverse.

Action is a class, not a score

A mutate/execute runtime isn't a "better" retriever than a read-only one — it's a different appliance doing a different job. If a grid renders read-only as a loss in a "can it write?" row, it has drawn the toaster failing at air-frying. Declare it; let the reader weigh it against their own charter.

Grounding is a real trade, not a defect

Embedding/RAG search trades line-exact fidelity for semantic recall — it finds "the thing that handles auth" when you don't know the symbol name. That's a genuine design axis. Calling it "less accurate" misses that it answers a question the grounded class structurally can't.