vs-token-safer

English · 한국어

A token-saving code layer for Claude Code on any codebase — TypeScript, JavaScript, Python, C#, C++, Go, and more. (Battle-tested down to a 26k-translation-unit Unreal Engine monorepo; bundles a game/build-log analyzer too.)

Your coding agent has a small context window. Your repo is large. vs-token-safer sits in between.

Ask where something is, what calls it, or even "how does the auth flow work?" when the name escapes you — and instead of pasting a wall of source into the chat, it replies with a short file:line list.

When your project builds normally, the answers are exact: it reads the same code index your editor relies on. When it doesn't, it still locates your functions and classes with nothing to set up. And when you've forgotten the name and only remember what the code does, it finds it from the vocabulary your own code already uses — no AI model, nothing uploaded. Markdown and config files work the same way: jump straight to one section by its heading instead of opening the whole file.

A companion plugin does the same for giant editor and build logs. None of it leaves your machine.

# Claude tries to grep code → the hook REWRITES it to the indexed query, in place:
$ grep -rn "createSession" src/
↻ [vs-token-safer] Rerouted → search_symbol "createSession"   # semantic, not a text match
  func createSession (in AuthService)  @ src/auth/session.ts:142   (+2 more)
  → ~120 tokens   (grep would have dumped thousands of lines)

# Editing that symbol? Name it — no Read-the-whole-file, no line counting:
$ replace_symbol_body symbol="createSession" body="…"        # preview; apply=true writes
  replace_symbol_body "createSession" — PREVIEW at src/auth/session.ts:142-160

Same flow on TypeScript, Python, C#, C++, Go and more (clangd · Roslyn · tsserver · pyright · tree-sitter). VTS_REWRITE=0 blocks instead of rewriting.

Why

• grep on a large repo — a TypeScript/Python monorepo, a C#/.NET solution, even a 26k-TU Unreal C++ tree — floods the context. The language-server index stays token-capped — ~97–99% smaller (benchmarks).

• Claude keeps reaching for grep. The hook doesn't just block it — it rewrites the command to the indexed query in place, so the search still runs and the flow never breaks.

• Edit by symbol, not by line. Replace/insert-around/delete a declaration by naming it — the index supplies the span, so you skip reading the whole file into context.

• You can't tell how much grep still slips through. vts discover reads your recent sessions and reports exactly which searches bypassed the index and what they cost.

• The language server runs headlessly — no editor open, unlike an IDE-proxy approach.

Related MCP server: codeix

Quickstart

# 1) Install (also auto-installs the gamedev-log-analyzer sibling)
/plugin marketplace add JSungMin/vs-token-safer
/plugin install vs-token-safer@vs-token-safer
/reload-plugins        # first run auto-installs the server deps (no manual npm)

# 2) Configure — detects the backend, asks for the project path, writes the config
/vs-token-safer:setup

Then restart the Claude Code session (the vs-search MCP server only starts on a fresh session). Verify the tools appear and that grep src/**/*.cpp is rerouted to the index. Prerequisites: Node ≥ 18 and a language server — clangd (C/C++) / Roslyn (C#) you install; JS/TS + Python auto-install. Details in Prerequisites below.

Want only the log analyzer? /plugin install gamedev-log-analyzer@vs-token-safer.

How it works

vs-token-safer isn't a search box — it's a precision ladder. You ask where something is, what calls it, or "how does the auth flow work?" when the name escapes you, and it answers at the highest precision it can reach, then tells you which rung the answer came from:

• EXACT — you know the name and the project builds → the official language server (clangd / Roslyn / tsserver / pyright), the semantic ground truth.

• SYNTACTIC — no toolchain set up → a tree-sitter parse (36 languages, bundled, no native build) still returns real declarations, not a grep.

• FUZZY — you only remember what the code does → a concept dictionary mined from the repo's own identifiers + comments (no AI model, nothing uploaded).

• SECTION — it's a doc or config, not code → Markdown / TOML / YAML / CSS / HTML addressed by heading.

The rungs aren't a one-time pick — they connect, and vts switches between them as it learns more. Start on FUZZY when you only know the intent; the moment concept_search surfaces a real name, vts climbs to EXACT to confirm it against the semantic index (and an exact search that misses drops back down to FUZZY). The hooks steer that hand-off in both directions, so "I don't know the name" turns into a precise, semantically-verified file:line instead of a dead end.

Every answer comes back capped to file:line (never source bodies) and carries a one-line completeness certificate naming the rung — so the model always knows whether it got the semantic truth or a fallback. On a 3-language, 150-file benchmark that's 87% fewer tokens than grep (~138× on a real Unreal Engine 5 tree). Underneath, four mechanisms make Claude actually use the ladder instead of reaching for grep:

Engine = official, glue = ours. clangd (LLVM) and Roslyn (Microsoft) do the analysis; this repo only writes the LSP↔MCP glue. No third-party MCP server runs over your source. Local-only, nothing uploaded.

Tools

All search/edit goes through an official language-server index — clangd (C/C++), Roslyn (C#/.NET), tsserver (JS/TS), pyright (Python) — and comes back as a compact, capped file:line list (never source bodies). MCP server vs-search; same tools as the vts CLI.

Search / navigate

Edit (symbol-level — name it, don't line-count) — preview by default, apply=true writes.

Docs & config too (structure tier). Point any of document_symbols / read_symbol / replace_symbol_body / insert_symbol / safe_delete at a Markdown / AsciiDoc / reST / TOML / INI / YAML / JSON / text file and the "symbol" is a section (heading, [section], or key): outline a 2000-line CLAUDE.md in ~30 lines, read or replace one ## Section by name without Reading the whole file. No language server, no new tools — same token-safer move, for documents.

Admin / meta — one MCP tool vts_admin {op, params} (folded to keep the per-session tool-definition cost small; the CLI keeps the bare subcommands):

e.g. vts_admin {op:"git", params:{argv:["status","-s"]}}. Or hand a whole "where is X / what calls Y / find file W" lookup to the code-locator subagent — it searches in its own context and returns only the file:line table.

Dashboard — vts serve. A local, interactive view of what vts knows + how much it saved: the savings trend, language mix, per-tool savings, and an interactive 3D graph (WebGL / Three.js) with two modes — the include graph (files sized by include fan-in) and an on-demand call graph (type a symbol → its transitive callers/callees, traced live through LSP callHierarchy — no persistent index; shows call counts per node/edge). Nodes are laid out on a spherical shell (so they spread out, not clump); drag/WASD to orbit, wheel/+-- to zoom, R to fit, hover for file:line. Symbol search has live autocomplete (/symbols); color by connected-component groups · repo (which repository each node is from, with a legend) · heat; click a node to drill into its group (Esc/Backspace to pop out); a focus/maximize toggle, a highlight filter, and a node/edge metrics overlay.

Easiest via the slash commands: /vs-token-safer:viz (open) and /vs-token-safer:viz-stop (close). Or the CLI:

vts serve --open     # → http://127.0.0.1:8731/  (launches the browser; --port N to change)
vts serve --stop     # stop it (or Ctrl-C the process)

It's 127.0.0.1-only and serves a fully self-contained page — CSS/JS inlined and Three.js vendored locally (server/vendor/, served same-origin, never a CDN), so nothing leaves the machine; it renders with the network unplugged. Same trust model as the rest of vts. Built on Node's stdlib http (no web-framework dependency), and it runs only when you invoke it — the MCP server never starts it, so the steady-state package stays a thin stdio client. The 3D graph caps at VTS_VIZ_MAX_NODES (200) for smoothness.

$ vts symbol --q createSession --projectPath ./app
3 symbol(s) matching "createSession" (backend: typescript, root: ./app):
func createSession (in AuthService)  @ app/src/auth/session.ts:142
method createSessionToken (in TokenStore)  @ app/src/auth/token.ts:88
func createSessionCookie  @ app/src/http/cookies.ts:31

✓ Saved ~4,200 tokens here (96.8% / 31× smaller than the raw index response).

The two plugins

vs-token-safer declares gamedev-log-analyzer as a dependency, so one install pulls in both. Used together: the log analyzer emits file:line per entry → hand it to goto_definition/find_references to open the code, without grepping or dumping the raw log. The handoff runs in reverse too — a code search aimed at a log (Logs/, .log/.jsonl) points you back at gamedev-log instead of an empty result.

Companion: drive vs-search with a local model

vts-local-orchestrator — a separate, optional companion that lets any local LLM (Ollama, full-GPU — model-agnostic, default gemma4:e4b, chosen by benchmark) drive these same vs-search tools. Claude delegates cheap, high-volume code-location to the free local model and receives only the compact file:line answer, so the raw search output never enters Claude's context.

• CLI (qvts) + a live web dashboard with a 3-way token-savings panel (this method vs CC-using-vs-search vs CC-using-grep), plus a delegation-routing skill for Claude Code.

• Token-savers: a persistent savings ledger (qvts --savings), a locate cache (zero-cost repeats), batch delegation (--batch), and terse repo-relative file:line output.

• Fully local — nothing transmitted off-machine, same as this plugin's charter.

Install: clone it next to this repo, npm install, then bash setup-macos.sh (Windows: setup.ps1). See that repo's README for the pipeline, the model benchmark, and the savings model.

Performance

A real A/B on a large Unreal Engine 5 project: finding one public engine symbol (FGameplayTag) via Bash grep-and-paste vs this plugin. No project source is reproduced, only aggregate counts; see BENCHMARK.md.

~99.3% fewer (~138×). grep returns the full text of every matching line and matches by text (comments, strings, unrelated identifiers); the plugin returns one file:line per semantic hit, capped. The mock-LSP eval (node eval/run.mjs, no toolchain) gates this on every commit: ~57,308 → ~1,549 tok = 97.3% (53/53 checks).

• Recall: the plugin returns the top N (cap), not every textual occurrence — the withheld tail is mostly comments/includes/substring noise. Need exhaustive? Raise maxResults, or use grep.

• Precision: grep matches every substring (Foo also hits FooBar); the index returns distinct semantic declarations.

So for navigation (a definition plus representative usages) the plugin is both more accurate and far cheaper. For an exhaustive occurrence audit, raise the cap or fall back to grep on purpose.

clangd indexes asynchronously, so the first search pays a one-time warm-up. vts handles this like an IDE: the MCP server pre-warms at boot (VTS_PREWARM, on when projectPath is set) and keeps the client cached for the session, so you pay it once. vts warmup builds the on-disk index up front (CLI/CI), and VTS_CLANGD_REMOTE points clangd at a shared prebuilt index server.

Ordering matters: clangd boosts the priority of files you open, so vts warms query-history-first, then what you're editing now (git status / p4 opened), then git-log recency, then include-centrality, then mtime. On a huge tree you can only warm a small slice, so this is what makes the warm window contain what you search for. Measured lift (node eval/bench-hitrate.mjs, 2,000 files):

The smaller the slice you can afford to warm, the bigger the win.

On a huge monorepo (e.g. a full Unreal Engine source tree, ~26k translation units) the cold index is the one real cost. Two opt-in levers cut it — both stay local, nothing is transmitted:

1. Scope — index a subtree, not the whole tree. See what you'd scope, then set it:

vts scope --projectPath /path/to/UE        # shows current scope, kept/total TUs, and top-level dirs to pick
vts setup --scope "MyGame,Plugins"         # persist it (or set VTS_SCOPE="MyGame,Plugins"); then reload/restart

clangd then indexes only the in-scope translation units (live UE5: MyGame → 3,377 of 26,488 TUs, 13%), and every backend's warm-up is scoped with it. No scope set = whole-tree behavior, unchanged.

2. Pre-index — build the index ahead of the first query.

vts preindex --projectPath /path/to/UE     # honors the scope above

With the full LLVM release installed (it bundles clangd-indexer next to clangd), this builds a monolithic static index offline and clangd loads it via --index-file — a local file, no server — so the first query is instant instead of waiting on the lazy background crawl. Without clangd-indexer it falls back to a warm pass (and tells you to install full LLVM). Override the binary with VTS_CLANGD_INDEXER_CMD.

3. Zero-setup tier — works before any toolchain, on any repo. No compile DB, no language server, no wait? vts still answers search_symbol from a tree-sitter parse (an official standard parser, 36 languages, bundled as wasm — no native build) — real declarations, not a usage grep, in the same token-capped file:line shape. Make it instant and shareable by committing an index:

vts index --projectPath /path/to/repo      # writes .vts-index/symbols.jsonl (commit it!)
vts index --status                          # show the current committed index

.vts-index/symbols.jsonl is plain, git-committable, and portable — commit it so teammates (and your own cold starts) get instant symbol search with zero setup. A language server, once it indexes, automatically supersedes it (the syntactic tier locates decls; the LSP adds reference/overload/type resolution on top). Benchmark (150-file symbol search): grep 4917 → tree-sitter 53 tokens = 98.9%, no toolchain.

Do existing users need to re-run setup? For default (whole-tree) behavior, no — just update the plugin and /reload-plugins. You only run vts setup --scope … (once) if you want to opt into scoping; the clangd-indexer path needs no vts setup at all (it's auto-detected — you just need full LLVM installed).

Each search records the tokens it saved vs forwarding the raw index response. Check it with /vs-token-safer:savings, vts savings (--graph/--daily/--history), or reset via vts savings-reset.

vs-token-safer savings (local, 1 search(es))
  total saved: ~4,200 tokens vs forwarding raw index responses
  raw → output: 4,340 → 140 tok (~31× smaller)
  est. value: ~$0.01 (@ $3/Mtok — set VTS_USD_PER_MTOK)

That's the caught side. vts discover scans recent sessions for searches that went around the index:

$ vts discover --since 1
  86 code search(es) bypassed vts (Grep×48, Glob×18, grep×12, find×8)
  catch-rate: ~770,333 tok caught (via vts) vs ~28,692 still bypassing → 96.4% routed through vts

(Searches the hook blocked don't count as bypasses.) discover is local and read-only — it reads transcript metadata and tool I/O sizes, never ships any of it anywhere. --learn feeds the files past searches hit into the warm-up set, so each session leaves the index warmer.

Prerequisites (details)

• Node.js ≥ 18 on PATH.

• C/C++ → clangd ≥ 22 (releases). The clangd 19.1.x bundled with Visual Studio deadlocks indexing real Unreal TUs in server mode; vts warns on an older one. Needs a compile_commands.json. Prefer the full LLVM release — it bundles clangd-indexer alongside clangd, which vts preindex uses for an instant static index (see Big trees: scope & pre-index).

• C#/.NET → a Roslyn LSP. Install the VS Code C# extension (ms-dotnettools.csharp) — vts auto-detects Microsoft.CodeAnalysis.LanguageServer and its runtime from the bundle. Fallback: dotnet tool install --global csharp-ls. Needs a .sln/.csproj.

• JS/TS → typescript-language-server, Python → pyright. Ship as plugin deps, install automatically on the first session (one-time ~50 MB; JS/TS wants Node 20+, skipped on 18).

• Mixed repo? A query that targets a file uses that file's own language backend — a .py/.ts inside a C++/C# (clangd/roslyn-rooted) tree gets pyright/typescript automatically, so vts works in a UE tree with a Python tooling dir without a manual backend=. This even overrides a pinned backend / VTS_BACKEND when they conflict: one global server serves every repo you touch, so a backend:"clangd" set for a C++ project never sends another repo's .js/.cs/.py to clangd (which would answer -32001 invalid AST). A query with no file target (e.g. search_symbol by name) keeps the pinned backend.

clangd needs a compile database:

• Unreal: <UE>/Engine/Build/BatchFiles/RunUBT … -mode=GenerateClangDatabase. If targets build with clang-cl, add -Compiler=VisualCpp or it fails clang-toolchain validation.

• CMake: -DCMAKE_EXPORT_COMPILE_COMMANDS=ON.

No compile DB yet? You still get answers — search_symbol falls back to a bounded literal text search, labeled as such, and the first result carries a one-time advisory. The one-command fix: vts_admin {op:"gen_compile_db"} (CLI vts gen-compile-db) assembles the exact UBT command (finds the .uproject, derives the <Name>Editor target, locates the engine, adds -Compiler=VisualCpp). Dry-run by default; apply=true runs UBT and parks the DB outside the source tree (~/.vs-token-safer/db/<project>, with clangd's .cache/ next to it, so git/p4 reconcile never see an artifact). inTree=true keeps the classic project-root layout, protected by a VCS-ignore guard.

Not published to npm — install vts from a clone:

git clone https://github.com/JSungMin/vs-token-safer
cd vs-token-safer/server && npm install && npm link   # provides `vts`
# or run directly: node /path/to/vs-token-safer/server/cli.js symbol --q SpawnActor --projectPath /path/to/proj

Configuration

Settings live in ~/.vs-token-safer/config.json (read at startup — /reload-plugins after changes). Configure via /vs-token-safer:setup (guided), vts_admin {op:"setup"} / {op:"config"}, or vts setup --projectPath <root> --backend clangd. Backend auto-detects from the root. Precedence: env (VTS_*) > config file > default.

Updating: Claude Code caches the marketplace, so new commits aren't auto-fetched:

/plugin marketplace update vs-token-safer
/plugin update vs-token-safer
/reload-plugins
# then RESTART the session — REQUIRED.

⚠️ A new version only takes full effect after a session restart. /reload-plugins updates hooks/commands/skills, but the running vs-search MCP server serves the old tool code until you quit and reopen. Version history: Releases.

Precedence: VTS_* env > ~/.vs-token-safer/config.json > default.

Status & safety

• clangd & Roslyn live-verified — search_symbol/find_references/goto_definition confirmed against real clangd (incl. a real Unreal 5.x game project end-to-end) and Microsoft.CodeAnalysis.LanguageServer. Needs clangd ≥ 22 and a correct compile DB.

• Local-only, nothing uploaded. The hook only inspects the command string (honors VTS_ENFORCE=0); the language server runs over stdio; the only outbound call is the first-run npm install of the MCP SDK. It writes only its config + a local savings ledger under ~/.vs-token-safer/. See SECURITY.md and PRIVACY.md.

• Savings/benchmark numbers are response-shaping (raw index → capped); savings vs grep are larger (BENCHMARK.md).

Contributing

Issues and PRs welcome — bug reports, new backends/engines, language mappings, docs. Keep PRs small, evidence-backed, and free of proprietary data (real paths/symbols/project IDs); add an eval/run.mjs guard for any new code path. See CONTRIBUTING.md. If this saved you tokens, a star helps others find it. ⭐

Acknowledgments

vs-token-safer stands on ideas from the open-source code-intelligence community. With gratitude to:

• codebase-memory-mcp (DeusData) — the tree-sitter tags.scm call-site approach, the multi-hop call-hierarchy (trace_path) shape, and content-hash-keyed caching. Our difference: we keep the official language server as the semantic source of truth and use tree-sitter only as a zero-setup syntactic tier below it — no reimplemented type-resolution layer, no persistent semantic DB; everything stays local and nothing is transmitted.

• Codeix (montanetech) — the idea of a plain, git-committable JSONL symbol index. Our difference: ours is a cold-start accelerator that a language server automatically supersedes once it has indexed.

• Code Context Engine (elara-labs) — the token-savings framing for AI code search. Our difference: no embeddings/vectors (so no nearest-but-wrong retrieval) — exact file:line, token-capped.

• Serena — symbol-level editing (replace_symbol_body / insert_symbol / safe_delete), here layered on the LSP with preview-by-default.

• The tree-sitter project and tree-sitter-wasms for the prebuilt grammars that power the syntactic tier.

Each of these made vs-token-safer better. Thank you. (Reuse here always keeps our charter: official engines do the analysis, output is token-capped file:line, and nothing leaves your machine.)

References & related work

The design choices here have academic backing, and the field is moving fast. The companion papers carry the full treatment — paper/vs-token-safer.tex and the fuzzy-retrieval follow-up paper/fuzzy-concept-dictionary.tex. Key references:

Foundations we build on

• HyperAgent — Phan et al., arXiv:2409.16299. Generalist software-engineering agents; motivates giving the model structured tools instead of raw grep.

• Codebase-Memory: Tree-Sitter Knowledge Graphs for LLM Code Exploration via MCP — arXiv:2603.27277. The closest sibling. Our difference: official-LSP ground truth + capped file:line, no persistent semantic DB, nothing transmitted.

• Dead-code lineage — Rapid Type Analysis (Bacon & Sweeney, OOPSLA'96) and demand-driven reachability (the Go deadcode model) underpin vts dce's preview-only call-graph reachability.

Related & subsequent work (2024–2026) — two of these are now migrated into vs-token-safer (marked Migrated).

• LARGER: Lexically Anchored Repository Graph Exploration and Retrieval — Hu et al., arXiv:2605.16352. Formalizes "lexical anchor → structural expansion" — the academic shape of our fuzzy→climb-to-exact ladder, also without embeddings. Migrated → concept_search now expands the import-graph neighbourhood only from high-confidence lexical anchors (a neighbour lifts a symbol only if its own match clears a fraction of the strongest one; VTS_CONCEPT_ANCHOR_MIN), so a weak/cross-cutting neighbour can't drag its imports up the ranking.

• Pseudo-relevance feedback (RM3) — Lavrenko & Croft, SIGIR 2001 (the classical embedding-free retrieval-feedback technique). Migrated → concept_search runs a second pass that mines expansion terms from the top results' own identifiers + comments and re-scores — bridging a synonym the query couldn't reach lexically ("warm" → the warmset's warming/dominant; "reachable" → the dead-code module's fixpoint/cascades). Drift-guarded (terms need a ≥2-result consensus, idf-ranked, capped) and the climb seed stays the pre-feedback exact match, so it widens recall without chasing the feedback. VTS_CONCEPT_PRF=0 reverts to single-shot.

• Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient? — Hsu, Yang & Lin, arXiv:2605.10848. Empirical case that lexical retrieval suffices inside the agent loop — backing for our no-embeddings charter.

• One Tool Is Enough: RL for Repository-Level LLM Agents (RepoNavigator) — Zhang et al., arXiv:2512.20957. A single jump-to-definition tool, RL-trained, beats large multi-tool agents — a strong recent endorsement of LSP navigation over grep.

• DCE-LLM: Dead Code Elimination with Large Language Models — arXiv:2506.11076. LLM-judged dead code. Our difference: vts dce stays preview-only static reachability — deterministic, no model judgment, with safe_delete as the disposal backstop.

• cAST: Structural Chunking via Abstract Syntax Tree — Zhang et al., arXiv:2506.15655. Tree-sitter chunking for code RAG — the neighbor of our SYNTACTIC tier. Migrated → read_symbol now cuts an over-budget body at a whole-child AST boundary (the end of a complete member/statement), never mid-statement, so the returned source stays syntactically whole (falls back to the plain line cap when tree-sitter is absent).

• RepoGraph (arXiv:2410.14684) and LocAgent (arXiv:2503.09089) — repository code-graph localization; we reach the same call/dependency structure via live LSP call-hierarchy, with no prebuilt graph and no embeddings.

License

MIT © 2026 JSungMin