English · 简体中文

A unified Code RAG toolkit — index once, three-mode output: pack / search / observe.

→ codingverse.github.io landing page

codingverse turns a code repository into a single SQLite index (symbols, call edges, chunks), then serves that index through three complementary outputs:

• pack — assemble a token-budgeted, layered context file for an LLM
• search — hybrid retrieval (BM25 + call graph, fused via RRF)
• observe — a six-board dashboard to see what the index actually contains

Everything is built on tree-sitter parsing and a local call graph — no embeddings, no external services, no API keys.

• Why
• Features
• Install
• Quickstart
• The three modes
• CLI reference
• MCP integration
• Dashboard
• Language support
• Architecture
• Development
• Contributing
• License

Feeding a whole repo to an LLM wastes tokens; feeding hand-picked files misses context. codingverse indexes the repo once, then lets you extract exactly the slice you need:

• Pack a token budget worth of the most important code, compressing the rest to skeletons/outlines instead of dropping it.
• Pack only what changed (--changed / --since) plus its call-graph impact radius.
• Pack only what a query matches (--query) plus its call-graph neighborhood.
• Search with a real call graph so results include callers/callees, not just text matches.

• Layered packing — every symbol is rendered at one of four fidelity levels (full / skeleton / outline / omit) to fit a token budget, ranked by PageRank over the call graph.
• Diff-scoped packing — cv pack --changed / --since <ref> packs changed files plus their reverse call-graph impact.
• Query-scoped packing — cv pack --query "<text>" packs search hits plus their bidirectional call-graph neighborhood.
• Hybrid search — BM25 lexical retrieval fused with real call-graph expansion (callers + callees) via Reciprocal Rank Fusion.
• Call-graph navigation — callers / callees / impact traversal over resolved edges.
• Continuous indexing — cv watch re-indexes on file changes; cv serve --watch keeps the dashboard live.
• Observation dashboard — six boards over a zero-build SPA served by node:http.
• MCP server — 7 tools over stdio for editor/agent integration.
• Multi-language — TypeScript, JavaScript, Python, Go, Rust, Java.
• Local-only — SQLite index, no embeddings, no network calls.

Requires Node ≥ 20 and pnpm. codingverse is a pnpm monorepo; build from source:

git clone <your-fork-url> codingverse
cd codingverse
pnpm install
pnpm -r build

The CLI entry point is packages/cli/dist/bin.js. Run it directly, or wire a shell alias:

alias cv="node /absolute/path/to/codingverse/packages/cli/dist/bin.js"

The rest of this README assumes cv resolves to that binary.

# 1. Build the index for a repo (symbols, call edges, chunks)
cv index /path/to/repo

# 2. Compute PageRank so packing/search can prioritize important symbols
cv rank /path/to/repo

# 3a. PACK — write a 32k-token context file
cv pack /path/to/repo --budget 32000 -o context.xml

# 3b. SEARCH — hybrid retrieval
cv search "token budget" /path/to/repo

# 3c. OBSERVE — open the dashboard at http://127.0.0.1:7331
cv serve /path/to/repo

# Keep the index hot while you work
cv watch /path/to/repo

Scoped packing for focused context:

# Only files changed vs HEAD + their impact radius
cv pack . --changed --budget 16000

# Only files changed since a ref
cv pack . --since main --budget 16000

# Only files a query matches + their call-graph neighborhood
cv pack . --query "retry backoff" --budget 16000

Full option details in docs/cli-reference.md.

The cv-mcp server (in @codingverse/mcp) exposes the engine over stdio via the Model Context Protocol, with 7 tools:

search · pack · expand · get_file · callers · callees · impact

Point an MCP-capable client at the built server binary (packages/mcp/dist/bin.js). Example client config:

{
  "mcpServers": {
    "codingverse": {
      "command": "node",
      "args": ["/absolute/path/to/codingverse/packages/mcp/dist/bin.js", "/path/to/repo"]
    }
  }
}

The repo must be indexed first (cv index).

cv serve starts a zero-build single-page app on 127.0.0.1:7331 (localhost-only by default). Six boards:

1. Overview — index stats, health, language breakdown
2. Token map — treemap of where the token budget goes
3. Code graph — interactive call graph with caller/callee highlighting
4. Retrieval — search inspector: BM25 vs call-graph paths, fused by RRF
5. Pack — live layered-pack preview with budget/strategy controls and full-output export
6. Sync — last index run and changed-file state

Use cv serve --watch to have the boards refresh as you edit code.

Adding a language is one tree-sitter tags query plus one registry entry — see docs/architecture.md.

codingverse is a pnpm monorepo of five packages:

The pipeline: ingest (walk + gitignore) → parse (tree-sitter → symbols + refs + chunks) → index (SQLite + FTS5, resolve refs into call edges) → assemble (layered pack / hybrid search). See docs/architecture.md for the full design.

pnpm install
pnpm -r build       # build all packages
pnpm -r test        # run the test suite (233 tests, centralized in core)
pnpm -r typecheck   # type-check without emitting

Tests live in @codingverse/core and cover parsing, indexing, resolution, ranking, packing, search fusion, scoped packing, watching, and multi-language extraction.

Contributions are welcome. See CONTRIBUTING.md for the development setup, project structure, and PR flow. Please also read the Code of Conduct.

MIT