codeindex

Self-contained, deterministic repo-indexing engine: file walking, language detection, symbol/import extraction (tree-sitter AST with a regex fallback), import resolution, a typed cross-file link-graph, and graph analytics — shipped as a single zero-dependency engine.mjs that consumer tools vendor (copy into their repo) instead of installing.

Designed for downstream tools — agent skills, CLIs, CI gates — that vendor the engine as a single file instead of taking an npm dependency.

What it does

• Walk a repo deterministically: ignore lists, binary/lockfile skips, a size cap, symlink-cycle guard. No file-count cap unless you ask for one (--max-files), and asking sets the capped flag — never a silent truncation.
• Scan every file into a FileRecord: classification, language, symbols, imports, headings, hashes — with an incremental cache fastpath. Extraction runs across worker threads by default (--workers, CODEINDEX_WORKERS); artifacts are byte-identical either way, and anything that would make a worker's result differ falls back to the single-threaded path.
• Extract symbols via tree-sitter (13 languages, when the wasm sidecar is present) or per-language regex rules (15 languages, always available).
• Resolve imports across languages: tsconfig paths, package exports, go.mod, Cargo, Java packages, PSR-4, C# namespaces.
• Build a typed link-graph: import / call / use / doc-link / mention edges at file and module level, plus Louvain communities, PageRank/ betweenness centrality, a tests→code map, and surprise-edge detection.
• Render byte-stable graph.json / symbols.json (two builds of an unchanged repo are byte-identical), plus a SCIP code-intelligence index (index.scip) via a hand-rolled zero-dependency protobuf encoder — validated by the official scip CLI (stats/lint).

Use as a library (the vendoring model)

Consumers commit scripts/engine.mjs + scripts/engine.d.mts (fetched at a pinned release tag) into src/vendor/ and import from it; their bundler inlines the engine so they still ship a single file:

import { buildIndexArtifacts, renderGraphJson } from "./vendor/engine.mjs";

const { scan, graph, symbols } = buildIndexArtifacts("/path/to/repo");

The AST tier is optional: without a grammars/ directory next to the bundle the engine silently uses its regex tier. Only tools that want AST precision also vendor scripts/grammars/ (~17 MiB of wasm).

Slim grammars (pull instead of vendor)

Consumers that want AST precision but not the ~17 MiB of vendored wasm can codeindex grammars pull the grammars once into a shared, per-machine cache (<XDG_CACHE_HOME|~/.cache>/codeindex/grammars/<ENGINE_VERSION>) instead:

codeindex grammars status   # active tier (adjacent/env/cache/none) + whether a pull is needed
codeindex grammars pull     # fetch the per-release grammars asset, sha256-verified, into the cache

Resolution is adjacent > env > cache > regex: a bundle-adjacent grammars/ still wins if present (offline setups are untouched), then CODEINDEX_GRAMMARS_DIR, then the pulled cache. pull fetches the official grammars-<version>.tar.gz release asset (its .sha256 sidecar is verified before anything is written) and extracts it atomically; the same wasm bytes produce byte-identical AST extraction from the cache as from a vendored dir. It is fully offline-safe: with no grammars resolvable anywhere — and after a failed or absent pull — the engine silently falls back to the regex tier exactly as it does today; a pull never throws into indexing.

Use from npm

For consumers who don't want to vendor the bundle, @maxgfr/codeindex also resolves as a regular package:

npm i @maxgfr/codeindex

import { scanRepo, ENGINE_VERSION } from "@maxgfr/codeindex";

const scan = scanRepo("/path/to/repo");

The CLI ships in the same package — see Use as a CLI below for the global install command. Consumer tools should still prefer vendoring: it keeps their own bundle single-file and pinned to an exact commit without an npm dependency.

Use as a CLI

brew install maxgfr/tap/codeindex        # or: npm i -g @maxgfr/codeindex

codeindex index   --repo . --out .codeindex   # graph + symbols + incremental cache
codeindex graph   --repo . > graph.json
codeindex scip    --repo . --out index.scip   # SCIP index (--out - for stdout)
codeindex callers --repo .                    # per-symbol caller index
codeindex grep    'pattern' --repo .

Docker

ghcr.io/maxgfr/codeindex ships the same zero-dependency bundle (engine.mjs

• cli.mjs + the AST grammars) with nothing else inside — just node and the files above, no npm install. Multi-arch (linux/amd64, linux/arm64), built and pushed on release. Mount the repo to index at /work:

docker run --rm -v "$PWD":/work ghcr.io/maxgfr/codeindex scan --repo /work
docker run --rm -v "$PWD":/work ghcr.io/maxgfr/codeindex index --repo /work --out /work/.codeindex

Pin by digest in CI or anywhere reproducibility matters, rather than a mutable tag:

docker run --rm -v "$PWD":/work ghcr.io/maxgfr/codeindex@sha256:... scan --repo /work

Runs as an MCP server over stdio the same way as the npm CLI (see Use as an MCP server below) — add -i so docker run keeps stdin open:

docker run -i --rm -v "$PWD":/work ghcr.io/maxgfr/codeindex mcp

Search

codeindex search "<query>" --repo . ranks files with keyless BM25 over symbol names, path segments, markdown headings and summaries. A query term that matches nothing in the corpus (zero document frequency) gets a deterministic trigram fuzzy fallback — typo tolerance without embeddings: the term is compared to the corpus vocabulary by character-trigram Dice similarity (threshold 0.6, top-3 candidates, contribution scaled by the Dice score so a near-miss always ranks below an exact hit). Terms that already match anything are never touched, so an existing query stays byte-identical. Enabled by default; disable with --no-fuzzy (CLI) or fuzzy: false (library/MCP SearchOptions.fuzzy); results carry an additive fuzzyTerms field when the fallback contributed.

Semantic search (deterministic static-embedding tier)

codeindex search "<query>" --repo . --semantic RRF-fuses lexical BM25 with a keyless, byte-deterministic embedding tier. It uses a static embedding model (a token → vector lookup table, no neural forward pass, no wasm): the pure-JS encoder tokenizes → mean-pools → L2-normalizes → int8-quantizes (round-half-to-even), and ranking is a pure integer dot product — so encode and the embeddings.bin artifact are byte-identical across builds and platforms.

It is opt-in by asset: with no model on disk the engine silently stays lexical, and --semantic without a model returns lexical results on exit 0 (a stderr note only). Models are never shipped in the package; a model is resolved from CODEINDEX_EMBED_DIR or <repo>/.codeindex/models/. Getting one is zero-config: codeindex embed pull fetches the official embed-model-v1 release asset, sha256-verified before anything is written.

codeindex embed pull   --repo .              # fetch the official model asset into
                                             # CODEINDEX_EMBED_DIR (or <repo>/.codeindex/models/); sha256-verified
codeindex embed status --repo .              # effective mode + reachability (JSON)
codeindex embed build  --repo . --out .codeindex   # write embeddings.bin
codeindex search "http client retry" --repo . --semantic

codeindex index also writes embeddings.bin next to graph.json when a model is present. Fusion reuses the engine's rrf helper (k=60); SCHEMA_VERSION is untouched (a dedicated EMBED_VERSION keys the sidecar).

Three embedding modes (precedence: endpoint > static > none)

mode	trigger	determinism
none	no model, no endpoint	— (pure lexical)
static	a model.json on disk	byte-deterministic (goldens)
endpoint	CODEINDEX_EMBED_ENDPOINT set	per image digest

The rich (endpoint) tier points the engine at a local containerized embedding server (all-MiniLM-L6-v2). The endpoint's float vectors flow through the same L2 + int8-quantize + integer-ranking pipeline as the static tier. Setting the env var is explicit intent, so it wins over a local model; an unreachable endpoint degrades to lexical (exit 0), not to the static model.

codeindex embed serve            # print the docker run one-liner (or --run it)
docker run -d -p 8756:8756 ghcr.io/maxgfr/codeindex-embed:latest
# reproducible: pin the digest → ghcr.io/maxgfr/codeindex-embed@sha256:<digest>
CODEINDEX_EMBED_ENDPOINT=http://localhost:8756 \
  codeindex search "auth token" --repo . --semantic

Full details incl. the HTTP protocol (build your own server): docs/SEMANTIC.md.

Use as an MCP server

codeindex mcp (or node scripts/cli.mjs mcp) serves the engine over stdio. Register it in Claude Code with:

claude mcp add codeindex -- codeindex mcp

26 tools, grouped by what they answer:

group	tools
orient	scan_summary, repo_map, graph, mermaid, workspaces
find	search, grep, find_symbol, symbols, symbols_overview
impact	find_references, callers, dead_code
risk	hotspots, churn, coupling, complexity, check_rules
edit (write)	replace_symbol_body, insert_after_symbol, insert_before_symbol
memory	write_memory, read_memory, list_memories, delete_memory (write except reads)
embeddings	embed_status

Pinning the server to one repository

Every tool takes a repo argument. A host that runs one server per workspace can pin it instead, so repo becomes optional on every tool — the pin is reflected in the advertised schema, not merely tolerated at call time:

codeindex mcp --repo /path/to/workspace

An explicit per-call repo still wins, so a pinned server can still answer about another checkout. --server-name <name> overrides the announced serverInfo.name for hosts that embed the server under their own identity.

Prime the index first and activation becomes a load, not a rebuild: codeindex index --repo <dir> --out <dir>/.codeindex. The first tool call deserializes those artifacts when the engine version, commit and artifact hashes all match. The same index also makes every CLI read command (search, symbols, graph, repomap, …) a lookup instead of a rebuild.

Protocol, and what it costs an agent

The server negotiates its protocol version: it answers with whatever revision the client asked for among 2024-11-05, 2025-03-26, 2025-06-18 and 2025-11-25, and otherwise with the newest. Fields a later revision introduced are only sent to clients that asked for it, so an older client sees exactly what it saw before.

From 2025-03-26 every tool carries behaviour annotations — readOnlyHint on the 21 read tools, destructiveHint/idempotentHint on the five that write — which is what lets a host auto-approve reads and confirm only writes. From 2025-06-18, the 15 tools whose result is always a JSON object also declare an outputSchema and return structuredContent, so a client can validate and type the result instead of re-parsing a string. The remaining tools return arrays, argument-dependent shapes or plain text, which cannot yield a conforming structured result without diverging from the text block — they are left unschema'd rather than described inaccurately.

Responses are capped (--max-response-bytes, default 1 MB). Under the cap nothing changes. Over it — where a whole-repo graph on a large monorepo runs to millions of tokens and no client can accept it — the response is replaced by a short notice naming the size, the artifact already on disk, and the narrower tool that answers the question. Most tools also take a limit/maxResults/ top/maxEdges argument to stay well under it.

engine.mjs is a pure side-effect-free library (safe for consumers to inline into their own CLIs); cli.mjs is the thin standalone CLI/MCP wrapper.

Command rewriting

codeindex rewrite '<command line>' maps an expensive tree-wide search onto its indexed equivalent, for agent harnesses that intercept shell commands (iterion's rewriters plugin kind, generalizing rtk):

$ codeindex rewrite 'grep -rn TODO src'
codeindex grep TODO --scope src

It prints the replacement and exits 0, or exits 1 with empty stdout when it has no opinion — run the original. The parser is deliberately conservative: any shell metacharacter (pipe, redirect, substitution, chaining), any unrecognized flag, a non-recursive grep, or more than one search path all refuse the rewrite. A refusal costs nothing; a wrong rewrite silently changes what the agent asked for.

Versioning

• ENGINE_VERSION — the release tag, embedded greppably in the bundle.
• SCHEMA_VERSION — the graph.json/symbols.json shape (currently 4). Consumers reject mismatched artifacts.
• EXTRACTOR_VERSION — the extraction output shape; incremental caches keyed on it are discarded wholesale when it bumps.

buildGraph/buildIndexArtifacts accept meta: { version, schemaVersion } so a consumer can stamp its own identity into artifacts it persists.

Benchmarks

Measured against universal-ctags, Serena (LSP over MCP) and Graphify with a reproducible harness (scripts/bench/); full methodology, fairness notes and all scenarios in BENCHMARKS.md.

Cold-index speed is not where this engine wins, and the table says so: a flat tags file is a smaller job and ctags finishes it first at every size. What a cold build buys is in the rows under it.

Metric	codeindex	Context
socialgouv/code-du-travail-numerique (2,823 files) — cold index	631 ms	vs ctags 330 ms, serena 7,695 ms, graphify 10,478 ms
vercel/next.js (27,952 files) — cold index	4,917 ms	vs ctags 3,357 ms; serena/graphify n/a — intractable at bench time
vercel/next.js — warm rerun / one file touched	1,234 ms / 2,489 ms	no competitor exposes an incremental reindex at all
Byte-identical rebuilds	7 / 7 repos	graphify's graph.json differed on all 6 it could be measured on; serena keeps no artifact
Install footprint	23.5 MB, zero runtime deps	serena 114.3 MB (+ language servers), graphify 140.1 MB
vercel/next.js — token ratio (measured)	390.3×	structured index vs raw grep, single-symbol lookup

Development

pnpm install
pnpm test          # unit + fixtures + compat + no-wasm gates
pnpm typecheck
pnpm build         # tsup → scripts/engine.mjs + scripts/engine.d.mts
pnpm check:build   # proves the committed bundle is byte-reproducible
pnpm test:e2e      # opt-in: pinned real-repo builds with ratchets

The compat suite pins golden bytes for the mini-repo fixture — the proof that extraction stays lossless across releases.

License

MIT