Virage Agent Claude

Monorepo for the Virage ecosystem — a Git-aware RAG pipeline that turns your codebase and docs into a searchable vector store. Pick your embedder and vector store, run one command, and keep the index in sync as code changes.

CI RAG Pipeline Release codecov Quality License: MIT

Packages

Core & CLI

Package	Version	Description
@vivantel/virage	npm	Rust CLI binary — primary install
@vivantel/virage-core	npm	Pipeline orchestrator, interfaces, config loading
@vivantel/virage-dashboard	npm	React web UI for pipeline monitoring

Chunkers

Package	Version	Description
@vivantel/virage-chunker-ce-ast	npm	Shared AST walker used by all CE native chunkers
@vivantel/virage-chunker-ce-ts	npm	TypeScript / JavaScript chunker (pure TS, no native binary)
@vivantel/virage-code-chunk-chunker	npm	AST-aware code chunking for TS, JS, Python, Go, Java

Embedders

Package	Version	Description
@vivantel/virage-embedder-openai	npm	OpenAI-compatible embedder (OpenAI, Azure, GitHub Models, Ollama)
@vivantel/virage-embedder-fastembed	npm	Fast local ONNX embeddings via FastEmbed
@vivantel/virage-embedder-transformers	npm	Local embeddings via @huggingface/transformers

Vector Stores

Package	Version	Description
@vivantel/virage-store-lancedb	npm	LanceDB vector store (embedded, file-based)
@vivantel/virage-store-qdrant	npm	Qdrant vector store (local and cloud)
@vivantel/virage-store-postgres	npm	PostgreSQL + pgvector vector store
@vivantel/virage-store-chromadb	npm	ChromaDB vector store (local or hosted)

Re-rankers

Package	Version	Description
@vivantel/virage-reranker-cross-encoder	npm	Local cross-encoder re-ranker (ONNX, no API key required)
@vivantel/virage-reranker-llm	npm	LLM-based re-ranker using the Anthropic API

Agent plugins

Package	Version	Description
@vivantel/virage-agent-core	npm	Base interfaces and utilities for agent plugins
@vivantel/virage-agent-claude	npm	Claude Code agent plugin
@vivantel/virage-agent-copilot	npm	GitHub Copilot agent plugin
@vivantel/virage-agent-codex	npm	OpenAI Codex agent plugin
@vivantel/virage-agent-antigravity	npm	Google Antigravity agent plugin
@vivantel/virage-skills	npm	AI agent skills for Claude Code, Copilot, and Codex

Utilities

Package	Version	Description
@vivantel/virage-git-isomorphic	npm	Pure-JS git source using isomorphic-git (no subprocess spawning)

Quick start

Install the CLI globally (requires Node.js 20+).

On Unix / macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/vivantel/virage/master/scripts/install.sh | bash

On Windows (PowerShell):

irm https://raw.githubusercontent.com/vivantel/virage/master/scripts/install.ps1 | iex

Or directly with npm:

npm install -g @vivantel/virage

Set up a project (interactive wizard generates virage.config.json):

cd my-project
virage init

Index your codebase:

virage index

Query the index:

virage query "how does authentication work?"

Or start the MCP stdio server for AI assistant integration:

virage serve

Configuration

All configuration lives in virage.config.json. Environment variables expand at runtime — use ${VAR} syntax for secrets (e.g. "apiKey": "${OPENAI_API_KEY}"). Example config:

{
  "providers": {
    "embedder": {
      "builtin": "onnx",
      "options": {
        "source": { "model": "Xenova/all-MiniLM-L6-v2", "cacheDir": ".virage/model-cache" },
        "dimensions": 384
      }
    },
    "vectorStore": {
      "builtin": "lancedb",
      "options": { "uri": ".virage/lancedb" }
    }
  },
  "fileSets": [
    {
      "name": "docs",
      "include": ["docs/**/*.md", "**/*.md"],
      "ignore": ["**/node_modules/**"],
      "chunkers": [{ "builtin": "md" }]
    },
    {
      "name": "code",
      "include": ["src/**/*.ts", "src/**/*.py", "src/**/*.go"],
      "chunkers": [{ "builtin": "lang" }]
    }
  ],
  "search": { "hybrid": true, "hybridAlpha": 0.6 },
  "pipeline": { "batchSize": 20 }
}

Run virage init to generate a config interactively. Run virage validate to check it. See docs/cli/config.md for the full config reference.

CLI commands

All commands have short aliases. Run virage --help to see the full list.

virage [options] [command]

Options:
  -v, --verbose   Increase verbosity (stackable: -v, -vv, -vvv…)
  --no-banner     Suppress the startup banner (also: VIRAGE_NO_BANNER=1)
  -h, --help      Display help

Commands:
  index     (i)     Run the indexing pipeline
  init              Generate virage.config.json interactively
  update    (up)    Update virage ecosystem packages and resync agent configs
  check     (c)     Validate embedder config matches the stored index
  validate  (val)   Validate config without running the pipeline
  dashboard (d)     Start the local monitoring dashboard
  query     (q)     Semantic search over the indexed knowledge base
  quality   (ql)    Quality system: self-assessment, retrieval eval, benchmarks
  report    (r)     Show observability report from pipeline runs
  chunks            Chunk analysis tools
  viz               Visualization tools
  store             Vector store diagnostics
  telemetry         Manage telemetry settings and data
  install-hooks (hooks)  Install git hooks for auto-indexing
  uninstall   (un)  Remove virage artefacts and optionally the global CLI

See docs/cli/ for per-command reference.

virage index flags:

Options:
  -c, --config <path>     Config file (default: virage.config.json)
  -f, --force             Force full rebuild
  --no-upload             Skip upload to vector store
  --dry-run               Show what would change without uploading
  --watch                 Re-run pipeline on file changes
  (Use VIRAGE_DIR env var to override the .virage/ directory path)

virage quality (ql) — 26-metric pipeline self-assessment, retrieval eval, and performance benchmarks. See docs/quality.md for the full command reference.

Chunkers

Chunkers split source files into embeddable chunks. PDF, Markdown, DOCX, LaTeX, and multi-language chunking are built into @vivantel/virage — no separate install needed. Configure chunkers per file set using the builtin: key:

{
  "fileSets": [
    {
      "name": "docs",
      "include": ["**/*.md"],
      "chunkers": [{ "builtin": "md" }]
    },
    {
      "name": "code",
      "include": ["src/**/*.ts", "src/**/*.py"],
      "chunkers": [{ "builtin": "lang" }]
    }
  ]
}

Plugin-based chunkers extend the built-in set:

Package	Language / format	Notes
virage-chunker-ce-ast	All (shared AST walker)	Base for TS/JS plugin chunkers
virage-chunker-ce-ts	TypeScript / JavaScript	Pure TS, no native binary
virage-code-chunk-chunker	TS, JS, Python, Go, Java	AST-aware, powered by code-chunk

Embedders

Key / Package	Requires API key	Notes
builtin: "onnx"	No	Built-in ORT inference; downloads from HuggingFace Hub; recommended default
virage-embedder-openai	Yes	OpenAI, Azure, GitHub Models, Ollama, any OpenAI-compatible endpoint
virage-embedder-fastembed	No	Fast local ONNX inference via FastEmbed
virage-embedder-transformers	No	HuggingFace Transformers.js; wider model selection

The embedder model and dimensions are tracked in the index. Changing either value automatically invalidates the cache and triggers a full re-embed on the next run.

Vector stores

Package	Infrastructure	Best for
virage-store-lancedb	None (file-based)	Local dev, CI, small projects
virage-store-postgres	PostgreSQL + pgvector	Production, complex SQL queries
virage-store-qdrant	Qdrant (Docker or cloud)	High-scale, distributed deployments
virage-store-chromadb	ChromaDB (Docker or hosted)	Simple hosted deployments

Re-rankers

Optional post-retrieval re-rankers re-score results for higher precision. Configured under providers.reranker in virage.config.json. When set, virage query applies reranking automatically — no extra flag needed.

Key / Package	Requires API key	Notes
builtin: "cross-encoder"	No	Built-in ORT cross-encoder; downloads from HuggingFace Hub
@vivantel/virage-reranker-llm	Yes (Anthropic)	LLM-based re-ranker using claude-haiku-4-5

Tuning

Fine-tune indexing performance via the options block in virage.config.json:

Option	Default	Effect
rateLimitMs	0	Milliseconds to wait between embedding API calls
batchSize	100	Chunks sent per embedding request
chunkConcurrency	CPU core count	Number of files chunked in parallel (I/O + AST parsing)
concurrency	1	Parallel embedding requests (for remote embedders only)

Local embedder models are cached in ~/.virage/models (overridable with VIRAGE_GLOBAL_DIR).

Use --force to discard the incremental cache and re-index everything from scratch.
Use -v / -vv / -vvv with virage index to increase log verbosity for debugging.

AI assistant integration

Agent plugins configure your coding assistant to use Virage for semantic search and context retrieval. Select one or more agents during virage init — the plugin is installed and configured automatically.

Agent	Plugin	What gets configured
Claude Code	@vivantel/virage-agent-claude	MCP server registration, slash commands, skills
GitHub Copilot	@vivantel/virage-agent-copilot	.github/copilot/ hooks and instructions
OpenAI Codex	@vivantel/virage-agent-codex	.codex/ hooks
Google Antigravity	@vivantel/virage-agent-antigravity	.antigravity/ hooks

Run virage update to resync agent configs after upgrading plugin packages.

Dashboard

Launch the web monitoring UI to inspect chunk distribution, embedding anomalies, pipeline status, search, and experiments:

virage dashboard        # start on port 3000
virage dashboard --port 8080 --verbose  # custom port + request logging

Use cases

See docs/USE_CASES.md for detailed scenarios:

• Onboarding new engineers with instant codebase search
• AI code review with full codebase context, not just the diff
• Keeping docs in sync with code via post-commit hooks
• Zero-cost local RAG for private or air-gapped codebases
• Multi-strategy indexing for mixed-content monorepos
• Measuring retrieval quality before and after config changes
• Cost-bounded CI indexing — only changed files are re-embedded
• Sharing project knowledge across Claude Code, Copilot, and Codex simultaneously

Roadmap

Planned features:

• Cross-file import graph indexing — so agents can follow call chains across files, not just within them
• Cost estimator (virage estimate) — projected token count and API cost before any embedding call
• PR diff mode — index only the files changed in a pull request, in an isolated namespace
• Semantic deduplication — skip near-duplicate chunks before embedding to reduce index bloat
• GitHub Actions integration — official vivantel/virage-action composite action for CI-driven indexing

Contributing

See CONTRIBUTING.md for setup instructions, commit conventions, and how to open a PR.

License

MIT