Architecture-Aware Code Intelligence for Massive Codebases.
Big Indexer (BGI) maps 100,000+ unit repositories by behavioral role, calculates blast-radius seams, and provides deterministic architectural context to AI coding agents via the Model Context Protocol (MCP).
$ pip install bigindexer && bgi mcp
Why Traditional Tools Break on Large Codebases
Vector search treats code like unstructured prose, while raw call-graphs explode into unreadable edge soup. BGI bridges the gap with behavioral tokenization and hard-bounded clustering.
Embeddings & RAG Fail
Chunk-based semantic search retrieves snippets that sound similar but completely misses structural call chains, runtime hierarchies, and cross-subsystem contracts.
Raw Call-Graphs Explode
Unscoped call graphs scale at O(N²), generating tens of thousands of noisy edges that collapse unrelated modules into unusable giant mega-clusters.
The BGI Architecture
BGI groups units by behavioral role (COV tokens), uses spectral masks to constrain edge generation, and caps cluster growth while emitting fuse-boundary signals.
The 3-Gate Deterministic Pipeline
Single-pass AST queries extract functional contracts, scoped spectral masks create behavioral edges, and hard-bounded community detection prevents over-merging.
Fingerprint Roles
Extracts units via Tree-sitter .scm queries. Assigns behavioral COV tokens (e.g. AUTHENTICATE, PERSIST, DISPATCH).
Scoped Matching
Forms Key-Lock edges between complementary units. Spectral masks restrict match scope to eliminate cross-file noise.
Bounded Clustering
DRS clustering merges tightly coupled units under a strict size cap. Refused merges emit explicit fuse-boundary signals.
MCP Context Server
Serves SQLite index to AI agents via MCP tools: task_fingerprint, behavioral_twins, and twin_context.
Head-to-Head Capability Matrix
| Capability | LSP / SCIP Index | Generic Call-Graph / Community | Big Indexer (BGI) |
|---|---|---|---|
| Symbol & Definition Lookup | Strong | Medium | Yes (Phase 6 Index) |
| Behavioral Token Model (COV) | No | No | Yes (Gate 1) |
| Hard-Bounded Cluster Growth | No | Rare (over-merges) | Yes (Max <2% repo) |
| First-Class Seam & Boundary Signal | No | No | Yes (fuse-graph.json) |
| Scope-Constrained Matching | File-only | No (unbounded O(N²)) | Yes (Spectral Masks) |
| Native Model Context Protocol (MCP) | Custom integrations | No | Yes (Official Registry) |
Two Commands to Full Codebase Intelligence
Run static architecture analysis locally or hook BGI directly into Claude Desktop, Cursor, or your CI/CD pipeline.
# 1. Install via pip
pip install bigindexer
# 2. Scan codebase & generate boundary artifacts
bgi scan /path/to/repo \
--lang auto \
--out bgi-graph.json \
--fuse-graph fuse-graph.json \
--routes routes.json
# 3. Launch MCP context server
bgi mcp --graph bgi-graph.json --fuse-graph fuse-graph.json// Prompt your AI coding agent
"Use MCP tool twin_context for task:
'Add rate-limiting middleware that validates tokens and logs latency.'
Return top behavioral twin, blast radius seam, and rubric checklist."
// BGI MCP returns:
Twin: auth_module.py::TokenValidator
Seam Risk: High (3 inbound links to dispatch_engine.py)
Checklist: [Handle TokenExpired, Assert metric_emit, Keep in Cluster 2]