Nexus-MCP

Hybrid search + code graph + semantic memory in a single local MCP server — under 350 MB RAM.

Nexus-MCP is a code intelligence server for the Model Context Protocol. It gives AI agents precise, token-efficient answers about your codebase without cloud dependencies: no API keys, no data egress, no subscriptions.

pip install nexus-mcp-ci
claude mcp add nexus-mcp-ci -- nexus-mcp-ci

The Problem It Solves

AI coding agents are token-inefficient by default. An agent trying to understand verify_credentials() typically:

1. Glob("src/**/*.py") → 120 files returned, agent reads the most likely 8 → ~12,000 tokens
2. Grep("verify_credentials") → 3 matches, agent reads surrounding context → ~4,000 tokens
3. Read("auth/middleware.py") → full 400-line file to understand callers → ~3,000 tokens

Total: ~19,000 tokens, 3+ tool calls, no graph relationships.

With Nexus-MCP:

1. explain("verify_credentials") → symbol definition + all callers + all callees + complexity metrics → ~1,500 tokens, 1 tool call

Or for discovery:

1. search("credential verification flow") → top-10 semantically relevant chunks across the codebase → ~2,000 tokens, 1 tool call

Estimated savings: 30–60% token reduction per coding session. The exact numbers depend on codebase size and task type — see the benchmarks table below.

Quickstart (60 seconds)

# 1. Install
pip install nexus-mcp-ci

# 2. Register with Claude Code
claude mcp add nexus-mcp-ci -- nexus-mcp-ci

# 3. Verify (in any Claude Code session)
# Claude will automatically use nexus-mcp-ci tools when CLAUDE.md instructs it

Then drop a CLAUDE.md in your project root:

## Code Navigation

Use nexus-mcp-ci tools before built-in file tools:
- Start sessions with `mcp__nexus-mcp__status`; run `index` if needed
- `search` before `Read/Grep`
- `explain` instead of reading a file to understand a symbol
- `impact` before any refactor

That's it. Claude will index your project on first use and use Nexus-MCP tools automatically.

How It Works

Indexing Pipeline (8 steps)

Source files
    │
    ├─ Step 1: Discover ──────── walk tree, filter by ext/size/.gitignore
    │
    ├─ Step 2: Parse symbols ─── tree-sitter (parallel ThreadPool)
    │           extracts: functions, classes, methods
    │           captures: name, signature, docstring, line_start/end, language
    │
    ├─ Step 3: Parse graph ────── ast-grep (sequential for consistency)
    │           extracts: call edges, import edges, inheritance edges
    │           output: UniversalGraph(nodes=[], edges=[])
    │
    ├─ Step 4: Transfer graph ── populate rustworkx PyDiGraph
    │           O(1) node lookup by name, Rust-backed traversal
    │
    ├─ Step 5: Chunk ──────────── Symbol → CodeChunk
    │           deterministic IDs: SHA256(file_path + symbol_name + line)
    │           avoids duplicate inserts on incremental reindex
    │
    ├─ Step 6: Embed ──────────── bge-small-en: 384-dim (default) or jina-code: 768-dim via ONNX
    │           lazy-loaded, unloaded after indexing (try/finally)
    │           GPU/MPS auto-detected; falls back to CPU
    │
    ├─ Step 7: Store ──────────── write to LanceDB `chunks` table (12-col PyArrow schema)
    │           rebuild native FTS (Tantivy) index after write
    │
    └─ Step 8: Cleanup ────────── unload model, persist metadata (mtimes for incremental)
                                  save rustworkx graph to SQLite (warm-start recovery)

Incremental reindex: mtime-based — only changed files are re-processed. Corrupt index detection triggers automatic full rebuild.

Search Pipeline

search("how does auth work")
         │
         ├─► vector_engine.search(query, n=30)  ← cosine similarity on 768-dim embeddings
         │                                         "auth" finds "verify_credentials", "token_check"
         │
         ├─► bm25_engine.search(query, n=30)    ← Tantivy FTS on same LanceDB table
         │                                         fast exact-keyword matching
         │
         ├─► graph_engine.boost(query, n=30)    ← structural relevance score
         │                                         hub symbols (high in/out degree) boosted
         │
         └─► fusion.merge(v_results, b_results, g_results)
                  │
                  │  Reciprocal Rank Fusion: score = Σ weight_i / (k + rank_i)
                  │  default weights: vector=0.5, bm25=0.3, graph=0.2
                  │
                  ├─► reranker.rerank(top_20)   ← FlashRank (optional, 4MB ONNX model, <10ms)
                  │
                  └─► token_budget.truncate()   ← summary / detailed / full
                           │
                           └─► Top-N chunks, scored, formatted

Technology Stack

Token Efficiency

Measured against equivalent agentic file-browsing workflows on a ~10,000-line Python codebase:

Typical session savings: 15,000–40,000 tokens (30–60%) compared to file-browsing agents.

Three Verbosity Levels

Every tool respects a verbosity parameter — agents request exactly the detail they need:

The 10 Tools

v2.0.0 breaking change: find_callers/find_callees/impact merged into graph, overview/architecture merged into map, and remember/recall/forget merged into memory — see CHANGELOG for the old→new mapping and ADR-017 for why. Fewer, richer tools route better under MCP Tool Search than many thin ones.

Discovery & Indexing

Search

Graph Analysis

Memory

Install

From PyPI (recommended)

pip install nexus-mcp-ci

# GPU (CUDA) support — adds ONNX CUDA execution provider
pip install nexus-mcp-ci[gpu]

# FlashRank reranker — adds ~4MB cross-encoder for better search quality
pip install nexus-mcp-ci[reranker]

# Both
pip install nexus-mcp-ci[gpu,reranker]

From Source

git clone https://github.com/jaggernaut007/Nexus-MCP.git
cd Nexus-MCP
./setup.sh           # creates venv, installs, verifies
# or
pip install -e ".[dev]"

Python 3.10–3.12 supported. Python 3.13+ is not yet supported by the current dependency stack, and the packaged Glama/Docker build uses Python 3.12 for compatibility. Optional: rg (ripgrep) for 100% search coverage fallback on unindexed files.

The optional jina-code model requires ONNX Runtime. If you see ONNX/Optimum errors:

pip install "sentence-transformers[onnx]" "optimum[onnxruntime]>=1.19.0"

The default bge-small-en model needs neither ONNX nor trust_remote_code.

MCP Client Setup

Claude Code

# Minimal
claude mcp add nexus-mcp-ci -- nexus-mcp-ci

# With the code-specific embedding model (requires trust_remote_code)
claude mcp add nexus-mcp-ci -e NEXUS_EMBEDDING_MODEL=jina-code -- nexus-mcp-ci

# GPU embeddings
claude mcp add nexus-mcp-ci -e NEXUS_EMBEDDING_DEVICE=cuda -- nexus-mcp-ci

# Virtualenv install — pass the full binary path
claude mcp add nexus-mcp-ci -- /path/to/.venv/bin/nexus-mcp-ci

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "nexus-mcp-ci": {
      "command": "nexus-mcp-ci",
      "args": [],
      "env": {
        "NEXUS_EMBEDDING_MODEL": "jina-code"
      }
    }
  }
}

Cursor / Windsurf / Cline / Any MCP Client

{
  "nexus-mcp-ci": {
    "command": "nexus-mcp-ci",
    "transport": "stdio"
  }
}

Agent Integration Patterns

CLAUDE.md boilerplate (drop into project root)

## Code Intelligence — nexus-mcp-ci

Every code task in this project MUST follow this workflow:

1. **Session start**: `mcp__nexus-mcp__status` → if not indexed, `mcp__nexus-mcp__index`
2. **Before any file read**: `mcp__nexus-mcp__search` to locate relevant code
3. **To understand a symbol**: `mcp__nexus-mcp__explain` (not Read)
4. **Before refactoring**: `mcp__nexus-mcp__impact` to assess blast radius
5. **For project orientation**: `mcp__nexus-mcp__overview` or `mcp__nexus-mcp__architecture`

Typical agent tool-call sequence

# Session start
status()               → "indexed: True, 8,412 chunks, 1,203 symbols, 87 MB"

# Code discovery
search("JWT token validation", mode="hybrid", n=10)
  → auth/jwt.py:42  validate_token()         score=0.94
  → auth/middleware.py:18  require_auth()    score=0.87
  → tests/test_auth.py:91  test_valid_jwt()  score=0.81

# Deep symbol understanding
explain("validate_token")
  → definition, docstring, params, complexity
  → callers: [require_auth, login_required, api_key_check]
  → callees: [decode_jwt, check_expiry, verify_signature]
  → quality: complexity=6, smells=[], maintainability=A

# Pre-refactor safety check
impact("validate_token")
  → direct callers: 3 symbols
  → transitive impact: 12 symbols across 4 files
  → high-risk: auth/middleware.py (5 dependents)

Multi-folder monorepo indexing

# Index multiple roots in one call — processed sequentially, shared engines
index(path="packages/api/src,packages/shared/src,packages/cli/src")

# Or use the paths parameter for additional roots
index(path="packages/api/src", paths="packages/shared/src,packages/cli/src")

Configuration

All settings via NEXUS_ environment variables:

Embedding Models

After changing model, re-index. Embeddings from different models are incompatible.

Comparison

vs. Other MCP Servers

vs. AI Code Tools

Development

git clone https://github.com/jaggernaut007/Nexus-MCP.git
cd Nexus-MCP
pip install -e ".[dev]"

pytest -v                    # 441 tests
pytest -m "not slow"         # skip performance benchmarks
pytest tests/test_search.py  # single module
ruff check .                 # lint

Project Structure

src/nexus_mcp/
├── server.py              # FastMCP entrypoint — 10 tools, input validation, graceful shutdown
├── config.py              # Settings (NEXUS_ env prefix)
├── state.py               # Global singleton SessionState
├── core/
│   ├── models.py          # Symbol, ParsedFile, CodebaseIndex, Memory
│   ├── graph_models.py    # UniversalNode, Relationship
│   ├── interfaces.py      # IParser, IEngine protocols
│   └── exceptions.py      # NexusException hierarchy
├── parsing/
│   ├── treesitter_parser.py   # Symbol extraction (parallel)
│   ├── astgrep_parser.py      # Structural graph extraction (sequential)
│   ├── language_registry.py   # 25+ language definitions
│   └── file_watcher.py        # Debounced watchdog for live reindex
├── engines/
│   ├── vector_engine.py   # LanceDB cosine similarity search
│   ├── bm25_engine.py     # LanceDB native FTS (Tantivy)
│   ├── graph_engine.py    # rustworkx PyDiGraph with RLock
│   ├── fusion.py          # Reciprocal Rank Fusion
│   └── reranker.py        # FlashRank (optional, graceful degradation)
├── indexing/
│   ├── pipeline.py        # 8-step indexing pipeline
│   ├── embedding_service.py   # ONNX Runtime, GPU/MPS auto-detect
│   ├── parallel_indexer.py    # ThreadPool over files
│   └── chunker.py         # Symbol → CodeChunk with deterministic IDs
├── memory/
│   └── memory_store.py    # LanceDB-backed memory, TTL, 6 types
├── analysis/
│   └── code_analyzer.py   # Cyclomatic/cognitive complexity, smells
├── security/
│   ├── permissions.py     # READ/MUTATE/WRITE tool categories
│   └── rate_limiter.py    # Token-bucket, per-tool, thread-safe
└── middleware/
    └── audit.py           # Structured audit logs, correlation IDs, field redaction

Adding a New Tool

1. Add the handler function to server.py decorated with @mcp.tool()
2. Add inline validation (_validate_* helpers in server.py) for any new input
3. Add permission category to security/permissions.py
4. Write tests in tests/
5. Update self_test/demo_mcp.py to exercise the tool

Adding a New Language

1. Add entry to parsing/language_registry.py with the tree-sitter grammar
2. Add structural patterns to parsing/astgrep_parser.py for call/import extraction
3. Add test fixtures in tests/fixtures/

Self-Test

Verify your installation exercises all 10 tools end-to-end:

python self_test/demo_mcp.py                   # built-in sample project
python self_test/demo_mcp.py /path/to/project  # your own codebase

Expected output: all 10 tools exercised with pass/fail per tool and a summary.

Known Limitations

• Sequential graph parsing: ast-grep runs sequentially (not parallel) to keep the call graph consistent. This is the main indexing bottleneck on large codebases.
• bge-small-en uses PyTorch: The lightweight model uses PyTorch instead of ONNX, so it doesn't benefit from the same ~50 MB footprint as jina-code.
• No incremental graph updates: Graph is rebuilt in full on incremental reindex (only vector/BM25 are incremental at the chunk level).
• No SSE transport: Only stdio transport is currently supported.
• Language coverage: 25+ languages, but structural relationship extraction (callers/callees) is most accurate for Python, TypeScript, JavaScript, Go, and Rust. Other languages may have partial graph edges.
• Static call graph only: find_callers/find_callees/impact are built from static parsing, not runtime tracing — dynamic dispatch, monkey-patching, and calls made through callbacks/closures/reflection won't show up as edges. Treat impact as a lower bound on blast radius in highly dynamic code.
• Auto-reindex has a detection lag: with the file watcher enabled (default), edits are picked up after a short debounce, and status()/search() run a throttled staleness check as a backstop — not an instant, per-call guarantee of freshness.

Architecture Decision Records

Key decisions are documented in docs/adr/:

Documentation

• Installation Guide — Prerequisites, client-specific setup, troubleshooting
• Architecture — Data flow, component design, memory budget analysis
• Usage Guide — Full tool reference with examples
• Developer Guide — Contributing, adding tools/engines/languages
• Research Notes — Library evaluations and technology deep-dives

Acknowledgments

Nexus-MCP consolidates two earlier open-source projects:

• CodeGrok MCP by rdondeti (Ravitez Dondeti, MIT) — Contributed the symbol extraction pipeline, embedding service, parallel indexer, core data models, and memory retrieval system.
• code-graph-mcp by entrepeneur4lyf — Contributed the ast-grep structural parser, rustworkx graph engine, complexity analysis, and relationship extraction.

Source files retain "Ported from" attribution in their module docstrings. See ADR-001 for the consolidation rationale.

License

PolyForm Noncommercial License 1.0.0. Free to use, copy, modify, and distribute for any noncommercial purpose. Commercial use requires a separate license — contact Shreyas Jagannath to inquire.

Versions published before 2.0.1 (0.1.0, 0.1.1, 2.0.0) remain available under their original MIT terms for anyone who obtained them under that license.