<div align="center"> <img src="logo/baner.png" alt="MSCodeBase Banner" width="100%"/>

🇬🇧 English • 🇷🇺 Русский • 🇨🇳 中文

MSCodebase Intelligence

AI-powered semantic code search for Zed IDE — deep code analysis MCP server

Features • Quick Start • Tools • Documentation • Installation • Architecture • Contributing • Security

Last updated: 2026-09-23

</div>

---

🎯 Positioning

MSCodeBase Intelligence is an MCP server for Zed IDE that gives AI assistants deep understanding of the entire codebase: semantic search, call graph, project memory, diagnostics.

This is not an LSP server or a replacement for the editor's built-in autocomplete. It's a "code intelligence" layer on top of the editor:

┌─────────────────────────────────────────────────────┐
│                      Zed IDE                        │
│  ┌───────────────────────────────────────────────┐  │
│  │        LSP (built-in autocomplete,            │  │
│  │        inline hints, diagnostics)             │  │
│  └───────────────────────────────────────────────┘  │
│                        │                            │
│                        ▼                            │
│  ┌───────────────────────────────────────────────┐  │
│  │  MSCodeBase (MCP server)                      │  │
│  │  · Semantic search across the codebase        │  │
│  │  · Call graph & impact analysis               │  │
│  │  · Project memory (ADR, tech debt)            │  │
│  │  · Self-diagnostics and self-healing          │  │
│  │  · 65 tools for AI assistant                  │  │
│  └───────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────┘

What you get

| Feature | MSCodeBase | Standard LSP (pyright/pylsp) | |---------|:----------:|:---------------------------:| | 🔍 Semantic search (BM25 + Vector + Reranker) | ✅ | ❌ | | 🧠 Call graph + impact analysis | ✅ | ❌ | | 🗃️ Project memory (ADR, known issues) | ✅ | ❌ | | 🏥 Self-diagnosis + self-healing | ✅ | ❌ | | 🔎 Cross-repo search | ✅ | ❌ | | 🤖 RAG answer generation (mode=ask) | ✅ | ❌ | | 🔬 Search explainability (per-stage score trace) | ✅ | ❌ | | 🏛️ Architecture drift detection (chain/circular/hub) | ✅ | ❌ | | ✅ Claim verification (agent fact-checking vs code) | ✅ | ❌ | | ✏️ Inline autocomplete | ❌ | ✅ | | 🏷️ Inlay hints | ❌ | ✅ |

LSP: Hybrid Rename Only

MSCodeBase uses LSP only for codebase(action="rename") — the LSP client (src/core/lsp_client.py) spawns pyright-langserver for precise cross-file rename, with graceful fallback to SymbolIndex (Tree-sitter) on timeout. All other functionality is implemented through 65 MCP tools.

The standalone LSP server (src/lsp_main.py) was removed — see LSP_WONTFIX.md.

Platforms

Designed and tested on Windows. macOS and Linux should work but have not been validated officially.

Languages

| Language | Parsing | Call Graph | Data Flow (ASSIGNED_FROM) | |---|---|---|---| | Python | ✅ | ✅ | ✅ | | TypeScript | ✅ | ✅ | ✅ | | TSX | ✅ | ✅ | ✅ | | Rust | ✅ | ✅ | ✅ | | Go | ✅ | ✅ | ✅ | | JavaScript | ✅ | ✅ | ✅ | | Java | ✅ | ✅ | ✅ | | C# | ✅ | ✅ | ✅ | | Ruby | ✅ | ✅ | ✅ | | PHP | ✅ | ✅ | ✅ | | Kotlin | ✅ | ✅ | ✅ | | Swift | ✅ | ✅ | ✅ | | C | ✅ | ✅ | ✅ | | C++ | ✅ | ✅ | ✅ | | Scala | ✅ | ✅ | ✅ | | Dart | ✅ | ✅ | ✅ | | Shell / Bash | ✅ | ✅ | ❌ (grammar without RHS-field) | | SQL | ✅ (context) | ❌ | ❌ | | YAML | ✅ (context) | ❌ | ❌ | | TOML | ✅ (context) | ❌ | ❌ | | HTML | ✅ (context) | ❌ | ❌ | | CSS | ✅ (context) | ❌ | ❌ | | HCL / Terraform | ✅ (context) | ❌ | ❌ |

✨ Features

| Feature | Description | |---------|-------------| | 🔍 Unified Search | search_code(query, mode, intent_hint) — single tool: fast/quality/deep/context/ask/auto | | 🧠 Intelligence Layer | 16 high-level intel_* tools: self-diagnostics, topology, memory, error prediction | | 🌐 Cross-repo Search | Search across multiple projects with @mention syntax | | 🌳 Call Graph | Full call graph: definition + callers + callees + impact analysis | | 🏗 Structural Search | 13 AST patterns (class_inheritance, async_function, decorator, etc.) | | 🔎 Context Search | Find similar code — paste a fragment, get semantic duplicates | | 🪣 Multi-Bucket RAG | Code/docs buckets, soft weighting, intent_hint (code/docs/auto) | | 🤖 mode=ask | RAG answer generation via phi-4 (server profile) | | 💾 LanceDB v2 | Vector DB with per-project isolation (incremental BM25 reindex) | | 🛡 Rate Limiting | DebounceBatch + CircuitBreaker — protection against VFS loops | | 🏥 Self-Diagnosis | get_health_report + index_health — full check and recovery | | 🧪 Clean Architecture | DI Container (14 services), 65 tools (32 core + 16 intel + 13 inline + 4 dev), ~1889 tests | | 🪟 Multi-Window | ProjectIndexerRegistry — isolated Indexer per project, LRU 5, ResourceMonitor throttle | | ✏️ Write Tools | codebase(action=...) — unified hub: rename, move, delete, replace, insert, ack | | ⚡ Meta-Patching | LanceDB move_chunks_metadata — file_path rename without re-embedding (50ms vs 5s) | | 🔗 Data Flow Graph | ASSIGNED_FROM edges track variable assignments. Unified Walker + Conditional Flow (if/for/while/try). 29 edge types in PropertyGraph. | | ⚙️ SYSTEM_PROFILE | light (sync) / server (async with phi-4) | | 🎯 MMR Diversification | Maximal Marginal Relevance (λ=0.6) after RRF — removes duplicates while preserving relevance. 0.3ms for 50 docs. | | 🧠 Auto Intent Detection | Keyword-based auto-detection of code/docs intent from query text. No manual intent_hint required. | | 📖 Extended Synonyms | 39 synonym groups (auth↔login, function↔method, cache↔buffer, etc.) — bridges the gap between user terminology and code. |

---

🚀 Quick Start

Install the mscodebase-intelligence extension in Zed, then:

cd <repo-root>
python install.py

# Quick sync (code only, no prompts):
python install.py --sync

# CI mode (no prompts, fail fast):
python install.py --yes

# Skip model downloads:
python install.py --skip-models

# Restart Zed (File → Quit → reopen)
# Verify: intel_get_runtime_status()

install.py does:

1. Copies 39+ source files to the extension directory
2. Installs Python dependencies
3. Downloads llama-server.exe + GGUF reranker model (bge-reranker-v2-m3). The embedder (multilingual-e5-small INT8) is an ONNX model downloaded separately.
4. Configures MCP in Zed's settings.json

See also: AI_INSTALLATION_PROMPT.md, docs/en/INSTALL.md

Providers

MCP auto-selects the best available provider (in priority order):

llama.cpp GGUF (native, preferred) → ONNX INT8 (in-process fallback) → LM Studio (if running) → BM25 only
   ~1.7 GB RAM (llama-server)          ~0.5 GB RAM                    ~6 GB RAM          no embeddings
   e5-small GGUF (384dim)              e5-small INT8 (384dim)         external API

Embedding runs via llama.cpp (llama-server.exe, preferred; ONNX in-process preload is canceled when llama.cpp is available). The reranker runs as a separate llama-server.exe process serving the BGE-M3 GGUF model. ONNX INT8 / LM Studio are fallback providers if llama.cpp is unavailable.

Benchmarks: docs/research/2026-07-10-final-benchmark.md

---

📚 Documentation Map

| Document | Description | Audience | Languages | |----------|-------------|----------|-----------| | docs/en/INSTALL.md | Installation, setup, uninstall | Users | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/ARCHITECTURE.md | Clean Architecture, Layers, DI | Developers | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/ARCHITECTURE_DEEP.md | Deep architecture: pipeline, lifecycle, comparison | Architects | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/SEARCH_PIPELINE.md | Search pipeline: BM25 → RRF → Reranker | Developers | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/GRACEFUL_DEGRADATION.md | 5 levels of graceful degradation (llama.cpp → ONNX → BM25) | DevOps | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/ARCHITECTURE_LAYERS.md | 10 runtime layers | Architects | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/FAQ.md | Frequently Asked Questions | All | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/TELEMETRY.md | Metrics, ETA, data collection | DevOps | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/investigations/ONNX_SESSION_REPORT.md | Full ONNX migration, 7 fixes, benchmarks | Support | 🇬🇧 | | docs/en/investigations/LSP_WONTFIX.md | LSP on Windows investigation (WONTFIX) | Support | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/ZED_WINDOWS_QUIRKS.md | Windows specifics, Restricted Mode | Windows users | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/CHANGELOG.md | Version history | All | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/CONTRIBUTING.md | How to contribute, PRs | Contributors | 🇬🇧 🇷🇺 🇨🇳 | | docs/en/SECURITY.md | Security policy, vulnerabilities | Security | 🇬🇧 🇷🇺 🇨🇳 | | AGENTS.md | AI Agent system rules | AI Agent | 🇬🇧 | | SECURITY.md | Security policy, reporting vulnerabilities | Security | 🇬🇧 | | CODE_OF_CONDUCT.md | Community standards | Contributors | 🇬🇧 | | CONTRIBUTING.md | How to contribute (root-level) | Contributors | 🇬🇧 | | KNOWN_ISSUES.md | Known issues & technical debt registry | All | 🇬🇧 |

All documents are cross-referenced. Available in 3 languages: English, Русский, 中文.

---

Research & Writeups

Deep-dives into specific technical findings from building this project:

• PageRank vs RAG on a Real Codebase: Corrected Numbers, and What I Almost Got Wrong Twice — comparing retrieval methods, with two rounds of honest self-correction
• I Asked One AI to Fact-Check Another AI's Audit of My Own Code
• The Silent Vector Contamination Bug: Why Your Concurrent Embeddings Might Be Lying to You

Recent results (Sept 2026)

• F5 4-arm unit-of-return: A 16.3% / B 34.4% / C 97.5% / D 0%; code 50% vs 6.3% → writeup
• NodeRAG refuted on the same bench: TF-IDF 80% vs graph BFS 70%
• D-ablation floor test (no-abstention): ~4% (2/48), majority 0/16 — prose B margin survives

---

🔧 MCP Tools (65 t