A structural code intelligence layer that transforms source code into a queryable knowledge graph for search, impact analysis, and AI-powered development.
Parse once. Query forever. Know exactly what breaks before it does.
📜 Licensed under PolyForm Noncommercial 1.0.0 — Source available for noncommercial use. • Need commercial use? Get a commercial license →
What is N3MO • Benchmarks • Architecture • Installation • MCP • Usage • Roadmap
N3MO is a symbol-centric code intelligence layer. Instead of scanning raw text, it parses your source code's ASTs, maps call graphs, and models dependencies in a queryable relational database — deterministically, with zero LLM calls at index time.
For engineering leaders and teams, N3MO acts as a structural insurance policy for your codebases.
🛡️ Eliminate Regression Risks— Utility functions are rarely refactored because developers fear unknown side effects. N3MO maps the transitive blast radius of any symbol to arbitrary depth, showing you exactly what will break before you make the edit.🏎️ Rapid Developer Onboarding— Instead of senior engineers spending hours explaining codebase flow to new hires, developers run one command to visualize complex call chains and parent-child dependencies interactively.🤖 AI-Agent Ready Infrastructure— Modern LLM agents (Cursor, Claude Desktop) are limited by context windows and text search. N3MO's native MCP server lets agents query the actual code graph, enabling fast, hallucination-free refactoring.⚡ No Embeddings, No Drift— N3MO is pure static analysis: Tree-sitter AST parsing into PostgreSQL. There's no vector index to keep in sync, no embedding cost per repo, and no semantic-similarity guesswork — every edge in the graph is an exact, verifiable relationship.
| Capability | Grep / Text Search | IDE "Find References" | N3MO Code Graph |
|---|---|---|---|
| Analysis Basis | |||
| Substring matching | AST-based, direct refs only | Relational knowledge graph | |
| Transitive Traversal | |||
| ❌ None | ❌ Manual, one level at a time | ⚡ Instant to arbitrary depth | |
| Blast Radius Mapping | |||
| ❌ None | ❌ Flat search-result list | 🎨 Interactive visual orbit map | |
| CI/CD Integration | |||
| ❌ None | ❌ Bound to IDE runtime | ⚙️ Dockerized CLI + CTE queries | |
| AI Agent Integration | |||
| ❌ Injected file chunks | 🤖 Native MCP server | ||
| Language Coverage | |||
| ✅ Any text file | ✅ 27 Tree-sitter grammars | ||
| Indexing Method | |||
| N/A | N/A | ✅ Deterministic AST parse — no embeddings, no LLM calls |
|
Developer: "Where does 'login' appear?"
Tool: grep -r "login" .
Result: 647 matches across 89 files
...now what?
|
Developer: "What breaks if I change login?"
Tool: n3mo impact "login"
Result: 3 direct callers → 5 ripple effects
Full blast radius in < 50ms
|
N3MO doesn't find text — it understands structure.It traces the actual call graph, not string matches.
Questions N3MO answers instantly:
| Question | How | |
|---|---|---|
| 🔎 | What functions and classes exist in this repo? | Full symbol index across 27 languages |
| 🎯 | Where is this symbol used — directly and transitively? | |
| Recursive CTE traversal to arbitrary depth | ||
| 💥 | What is the blast radius of changing this function? | |
| Interactive orbit map with depth slider | ||
| 🕸️ | How do these components actually connect? | Call graph + parent-child hierarchy |
| 🤖 | Can my AI agent understand this codebase structurally? | Native MCP server for Cursor / Claude |
All benchmarks measured on Intel i5-13450HX, 24 GB RAM, NVMe SSD.
Django is the primary benchmark target: 3,021 files, ~43K symbols, ~181K calls.
Django Index Time (minutes)
═══════════════════════════════════════════════════════════════
v0.3 Baseline ██████████████████████████████████████████████ 23 min 1×
SPLIT_PART Fix ██████████████████████ 11 min 2×
Batch Inserts █████████ 5 min 4.6×
+ Multiprocessing ████ 2.5 min 9× 🚀
═══════════════════════════════════════════════════════════════
| Optimization | Index Time | Speedup | What Changed |
|---|---|---|---|
| v0.3 baseline | 23 min | 1× | Per-symbol DB inserts, naive call resolution |
| + SPLIT_PART query fix | 11 min | 2× | Eliminated redundant string splitting in call resolution |
| + Batch inserts | 5 min | 4.6× | Symbols, imports, and calls batched per file (1 transaction) |
| + Multiprocessing | ~2.5 min | ~9× | |
ProcessPoolExecutor distributes AST parsing across cores |
✅ All results are real measurements on the
[Django]repository. Multiprocessing gains scale with core count.
Tested on TensorFlow — a 36,000-file, multi-language (C++/Python) monorepo.
| Metric | Result |
|---|---|
| Repo size (total files) | |
| ~36,000 | |
| Files processed & indexed | |
| 14,611 (after filtering tests, configs, and non-source files) | |
| Total symbols extracted | |
| 79,523 | |
| Total call edges extracted | |
| 480,851 | |
| Full index time (cold start) | |
| 14.06 minutes | |
| Peak memory (Docker container) | |
| 185 MB RAM | |
| CPU utilization | |
| ~5% |
N3MO scales from a 3K-file pure-Python repo (Django) to a 36K-file multi-language enterprise monorepo (TensorFlow) — roughly a
5× larger indexing job at near-linear throughput, without significant resource overhead. Symbol/edge-per-file and incremental (warm) re-index numbers for TensorFlow are being finalized in the full benchmark report.
Tested on ScanCode Toolkit — ~600K lines of Python.
| Metric | Result |
|---|---|
| Lines of code | |
| ~600,000 | |
| Full index time | |
| ~3 minutes | |
| Processing mode | |
| Single-threaded (v0.3) |
N3MO uses SHA-256 file hashing to skip unchanged files on subsequent runs.
| Scenario | Time | Notes |
|---|---|---|
| Full index (first run) | ||
| Baseline | All files parsed and inserted | |
| No changes (re-run) | ||
| < 1 second | ||
| Hash comparison only, zero DB writes | ||
| 1 file modified | ||
| < 2 seconds | ||
| Only the changed file is re-parsed and upserted |
These results are from the built-in benchmark script on a 20-file synthetic repository. Real-world incremental performance is proportional to the number of changed files, not the total repository size.
Impact analysis uses PostgreSQL recursive CTEs with cycle guards. Query times are independent of repository size — they depend only on the size of the result subgraph.
| Query Type | Typical Latency |
|---|---|
| Direct callers of a symbol | < 10 ms |
| Full blast radius (depth ≤ 5) | < 50 ms |
| Complete graph traversal | < 200 ms |
python benchmarks/benchmark_indexing.py
N3MO builds a symbol-centric knowledge graph stored in PostgreSQL:
graph TD
A["📄 Source Code"] -->|Tree-sitter| B["🌳 AST Parser"]
B --> C["🔍 Symbol Extractor"]
D["🔄 Git Hooks"] -->|post-commit| A
C --> E[("🗄️ PostgreSQL<br/>Projects · Symbols · Calls<br/>Imports · Files")]
E --> F["💥 Impact Analysis"]
E --> G["📞 Call Graph"]
E --> H["📊 Dependency Graph"]
F --> I["🎨 Visualizer"]
G --> I
H --> I
F --> J["🤖 MCP Server"]
style A fill:#6c63ff,stroke:#4a3fbf,color:#fff
style B fill:#7c74ff,stroke:#4a3fbf,color:#fff
style C fill:#7c74ff,stroke:#4a3fbf,color:#fff
style D fill:#ffd93d,stroke:#d4b800,color:#1a202c
style E fill:#ff6b6b,stroke:#c53030,color:#fff,stroke-width:3px
style F fill:#45b7d1,stroke:#2c8ea8,color:#1a202c
style G fill:#45b7d1,stroke:#2c8ea8,color:#1a202c
style H fill:#45b7d1,stroke:#2c8ea8,color:#1a202c
style I fill:#9ae6b4,stroke:#2f855a,color:#1a202c
style J fill:#ffd93d,stroke:#d4b800,color:#1a202c
sequenceDiagram
participant User as User / CI
participant CLI as N3MO CLI
participant DB as PostgreSQL (Docker)
participant Viz as Graph Visualizer
rect rgb(26, 27, 46)
Note over User, DB: Indexing Flow (Local CLI)
User->>CLI: n3mo index
CLI->>DB: Start PostgreSQL container (if not running)
CLI->>CLI: Walk file tree (SHA-256 hash checks)
CLI->>CLI: Parse AST (Tree-sitter, multiprocessing)
CLI->>DB: Batch insert symbols, calls, imports
CLI->>DB: Resolve imports & call links
DB-->>CLI: Success
CLI-->>User: Complete summary
end
rect rgb(26, 27, 46)
Note over User, Viz: Query & Visualization Flow
User->>CLI: n3mo impact "symbol" --graph
CLI->>DB: Recursive CTE traversal (depth & file filters)
DB-->>CLI: Blast radius subgraph
CLI->>Viz: Generate orbital vis.js HTML
CLI->>User: Launch local web server & open browser
end
erDiagram
PROJECT ||--o{ SYMBOL : contains
PROJECT ||--o{ CALL : tracks
PROJECT ||--o{ IMPORT : tracks
PROJECT ||--o{ FILE : indexes
SYMBOL ||--o{ CALL : "source of"
SYMBOL ||--o{ CALL : "resolved to"
SYMBOL ||--o{ SYMBOL : "parent of"
PROJECT {
uuid id PK
text name
text repo_url
timestamp created_at
}
SYMBOL {
uuid id PK
uuid project_id FK
text name
text file_path
text kind "function|class|method"
text signature
int start_line
int end_line
uuid parent_id FK
}
CALL {
uuid id PK
uuid project_id FK
uuid source_symbol_id FK
text call_name
int line_number
uuid resolved_symbol_id FK
}
IMPORT {
uuid id PK
uuid project_id FK
text file_path
text module
text name
text alias
uuid resolved_symbol_id FK
}
FILE {
uuid project_id FK
text file_path PK
text sha256
}
Multi-language support— 27 Tree-sitter grammars supported (dynamically loaded); actively benchmarked on 10 major languages including Python, JS/TS, Go, Java, and C/C++Parallel AST ingestion—ProcessPoolExecutor
distributes CPU-bound parsing across all available coresIncremental re-indexing— SHA-256 file hashing skips unchanged files automatically** Idempotent operations**— re-indexing updates existing data without duplication** Smart exclusions**— case-insensitive directory filters and camelCase-aware filename checks prevent false positives (e.g. allowscontest.py
while skippingtest_*.py
)
Symbol extraction— functions, classes, methods with full file path + line context** Hierarchical modeling**— parent-child relationships (Module → Class → Method)** Call graph construction**— who calls whom, resolved at ingestion time** Scope-aware resolution**— class scope > local file > imports > qualified dot paths > global** Blast radius analysis**— recursive CTE traversal to arbitrary depth with cycle guards
Connection pooling—ThreadedConnectionPool
eliminates per-symbol DB round tripsBatch inserts— symbols, imports, and calls batched per file in single transactions** Optimized queries**—SPLIT_PART
fix delivered a 2× speedup on call resolution
Interactive graph— vis.js orbit map with click-to-inspect nodes, sidebar, and depth slider** Dark mode**— toggleable canvas dark mode with real-time node/edge updates, persisted inlocalStorage
Premium styling— sleek interactive dashboard landing page UI and graph visualizer styled withBricolage Grotesque
,Inter
, andJetBrains Mono
typographyprofile — system instructions to configure Claude as an impact-aware coding agentSKILL.mdNative MCP server— first-class integration with Cursor, Claude Desktop, and Windsurf** Git hooks**— automatic re-indexing on every commit** CI pipeline**— GitHub Actions with linting (ruff
), type checking (mypy
), andpytest
Tree-sitter parsing supported for 27 languages. Deep semantic call graph mapping currently optimized for Python, JS/TS, and Java.
Install N3MO directly from PyPI:
pip install n3mo
n3mo setup
Alternatively, for contributors running in editable mode:
git clone https://github.com/RajX-dev/N3MO.git
cd N3MO
pip install -e .
n3mo setup
N3MO includes a native MCP server that exposes repository analysis and graph traversal tools to LLM agents (like Claude, Cursor, or Windsurf).
n3mo mcp install
This registers N3MO and sets up the paths automatically. Restart Claude Desktop and you're ready!
To configure Claude to run N3MO impact queries proactively before changing code in the editor, import or copy-paste the custom instructions from the ** SKILL.md** profile.
- Go to
Settings → Models → MCP. - Click
+ Add New MCP Server. - Set the configuration details:
Name:
n3mo
Type:command
Command:n3mo mcp start
(oruvx n3mo mcp start
to run directly)Environment Variables:TARGET_CODE_DIR=/absolute/path/to/your/active/workspace
- Click Save, and Cursor will instantly be able to index and query your workspace blast radius.
| Tool | Description |
|---|---|
n3mo_index |
|
| Ingests and indexes the codebase | |
n3mo_search_symbol |
|
| Locates the definition of a symbol across the workspace (file path, line number) | |
n3mo_get_dependencies |
|
| Finds all external symbols that a given symbol calls (forward-dependency graph) | |
n3mo_get_file_symbols |
|
| Lists all classes and functions defined inside a specific file | |
n3mo_get_blast_radius |
|
| Traces the transitive impact/call graph of a code symbol |
For team collaboration and automated pull-request analysis in your CI/CD pipeline, visit ** n3mo.shop** to get started with the GitHub Webhook integration.
N3MO is free under the PolyForm Noncommercial 1.0.0 License for local usage and single-developer MCP integrations.
100% Free & Local— CLI queries, local MCP integrations, and the visualizer with zero limits.** Enterprise Licensing**— for large-scale organization deployments or commercial licensing terms, reach out to the author.
cd /path/to/your/project
n3mo index
What gets indexed:
- ✅ Source files in all 27 supported languages
- ❌ Virtual environments (
venv/
,.venv/
) - ❌ Dependencies (
node_modules/
,site-packages/
) - ❌ Build artifacts (
.git/
,__pycache__/
,dist/
) - ❌ Test / fixture directories (
tests/
,mocks/
,specs/
)
Example terminal output:
◈ IMPACT ANALYSIS
──────────────────────────────────────────────────────────────────
Target: authenticate_user
──────────────────────────────────────────────────────────────────
◉ Direct Callers (3 symbols)
▸ login_endpoint api/auth.py:12
▸ refresh_token api/token.py:23
▸ validate_session middleware/auth.py:89
◎ Ripple Effects (5 symbols)
╰─▸ POST /login routes.py:67
╰─▸ admin_login admin/views.py:34
╰─▸ require_auth decorators.py:12
╰─▸ dashboard_view views/dashboard.py:8
╰─▸ settings_view views/settings.py:22
──────────────────────────────────────────────────────────────────
Total impacted: 8 references │ depth ≤ 3
php
graph LR
A[main.py] --> B[auth.py::login]
A --> C[db.py::connect]
B --> D[utils.py::hash_password]
B --> E[models.py::User]
C --> F[config.py::DB_URI]
style A fill:#ff6b6b,stroke:#c92a2a,stroke-width:2px,color:#fff
style B fill:#4ecdc4,stroke:#0ca89e,stroke-width:2px,color:#000
style C fill:#45b7d1,stroke:#1098ad,stroke-width:2px,color:#000
style D fill:#96ceb4,stroke:#63b598,stroke-width:2px,color:#000
style E fill:#ffd93d,stroke:#f5c200,stroke-width:2px,color:#000
style F fill:#e0e0e0,stroke:#a0a0a0,stroke-width:2px,color:#000
All four development phases have been completed. N3MO is stable and actively maintained.
| Phase | Component | Status |
|---|---|---|
| Phase 1 — Foundations | ||
| Docker setup | ✅ Complete | |
| Database schema | ✅ Complete | |
| Tree-sitter integration | ✅ Complete | |
| Symbol + call extraction | ✅ Complete | |
| Blast radius (recursive CTE) | ✅ Complete | |
| Interactive visualizer | ✅ Complete | |
| Phase 2 — Performance | ||
| Connection pooling | ✅ Complete | |
| Batch DB operations (symbols/imports/calls) | ✅ Complete | |
| SPLIT_PART query optimization | ✅ Complete | |
--file / --depth CLI flags |
||
| ✅ Complete | ||
| Interactive depth slider | ✅ Complete | |
| Phase 3 — Correctness & Scaling | ||
| Incremental re-index (file hashing) | ✅ Complete | |
| Multiprocessing (AST parsing) | ✅ Complete | |
| Scope-aware call resolution | ✅ Complete | |
| CTE cycle guard | ✅ Complete | |
| Full type annotations + mypy | ✅ Complete | |
| pytest suite + CI | ✅ Complete | |
| Multi-language support (27 languages) | ✅ Complete | |
| Phase 4 — Distribution | ||
| MCP server (Cursor / Claude / Windsurf) | ✅ Complete | |
| Real-time git-hook indexing | ✅ Complete |
Phase 1: Foundations ✅ Complete
- Docker environment (PostgreSQL)
- Database schema — Projects, Symbols, Calls, Imports tables
- Tree-sitter parser integration
- Symbol extractor with full AST traversal
- Idempotent upsert logic
- Blast radius via recursive CTE
- Interactive vis.js visualizer
Phase 2: Performance ✅ Complete
psycopg2.pool.ThreadedConnectionPool
— replace per-call connections -
execute_values()
batch inserts for symbols, imports, and calls — 1 transaction per file - SPLIT_PART query optimization for call resolution #
--file
and--depth
CLI flags for targeted impact analysis - Interactive depth slider in visualizer
Results: Django (3,021 files, ~43K symbols, ~181K calls) — 23min → 2.5min (9× faster)
Phase 3: Correctness + Scaling ✅ Complete
- SHA-256 file hashing for incremental re-index
ProcessPoolExecutor
for parallel AST parsing - Scope-aware call resolution using imports table
- CTE cycle guard (visited node tracking)
- Full type annotations,
mypy
clean checking in CI - pytest unit + integration test suite
- GitHub Actions CI pipeline
- Multi-language support (27 languages)
Phase 4: Distribution ✅ Complete
- MCP server — N3MO as a tool for Cursor, Claude Code, Windsurf
- Real-time incremental indexing via git hooks
1. Structure before semantics Map the code skeleton (AST) before adding AI analysis. A correct graph is worth more than a smart but wrong one.
2. Database as source of truth All state lives in PostgreSQL, eliminating in-memory complexity and enabling graph queries that application-level traversal cannot match.
3. Correctness over speed The parser must handle syntax errors gracefully without corrupting the graph. A fast indexer that silently drops symbols is worse than a slow one that gets everything right.
4. Idempotent operations Re-running ingestion produces identical results, enabling safe incremental updates and CI/CD integration.
Contributions are welcome! Please read the ** CONTRIBUTING.md** guide to get started with setting up the project, coding standards, and running checks locally.
pip install -e ".[dev]"
ruff check n3mo/
mypy n3mo/
pytest tests/
Licensed under the PolyForm Noncommercial 1.0.0 License.
- ✅ Free for personal projects, academic research, and hobby tools
- ✅ Source available — view, modify, and distribute for noncommercial purposes ⚠️ Noncommercial — you may not use it for commercial purposes⚠️ Restrictions apply on offering it as a service
For commercial deployments or proprietary modifications, contact for licensing options.
See LICENSE for full legal details.
Raj Shekhar — Delhi Technological University
— for robust, incremental, error-tolerant parsingTree-sitter— for making recursive graph queries possible without a graph databasePostgreSQL— for reproducible, single-command environmentsDocker— for the interactive graph visualizationvis.js— for the high-performance REST layerFastAPI