loading…
Весь каталог
LocalNest
БесплатноA local-first MCP server that provides AI agents with safe codebase access through file discovery, hybrid lexical-semantic search, and project introspection. It
автор: wmt-mobile
Описание
A local-first MCP server that provides AI agents with safe codebase access through file discovery, hybrid lexical-semantic search, and project introspection. It features durable local memory and semantic indexing while keeping all data and processing entirely on your local machine.
README
LocalNest MCP
Code Intelligence + Knowledge Graph + AI Memory — One Local MCP Server
The only MCP server that unifies semantic code search, a temporal knowledge graph, and persistent AI memory in a single local-first package.
74 tools. Zero cloud. Pure SQLite.
Why LocalNest?
Every other MCP server forces you to choose: memory or code intelligence. Never both.
LocalNest is the first to combine all three pillars into one server that runs entirely on your machine:
| Pillar | What it does | Why it matters |
|---|---|---|
| Code Intelligence | Hybrid BM25+vector search, AST-aware chunking, symbol finding (defs/usages/callers) | Your AI understands code structure, not just text |
| Knowledge Graph | Temporal entity-triple store with multi-hop traversal and as_of time-travel queries |
Architectural decisions, dependencies, and facts — versioned over time |
| Persistent Memory | Cross-session recall, semantic dedup, agent-scoped isolation, conversation ingestion | Your AI remembers what you taught it — forever |
How LocalNest Compares
No other MCP server covers all three pillars. Here's how the landscape breaks down:
vs Memory-Only Servers
| LocalNest | Mem0 | Basic Memory | MCP Memory Service | AgentMemory | |
|---|---|---|---|---|---|
| Persistent AI memory | Yes | Yes | Yes | Yes | Yes |
| Knowledge graph | Yes | No | No | No | No |
| Semantic code search | Yes | No | No | No | No |
| Symbol finding (defs/usages) | Yes | No | No | No | No |
| AST-aware chunking | Yes | No | No | No | No |
| Local-first / no cloud | Yes | Hybrid | Yes | Yes | Yes |
| MCP tools | 74 | 8 | ~10 | 24 | 43 |
Mem0 has 41k stars and $24M in funding — but it's memory-only with no code intelligence. Basic Memory integrates with Obsidian but can't search code. AgentMemory has auto-capture hooks but zero code features.
vs Code Intelligence Servers
| LocalNest | GitNexus | claude-context | codebase-memory-mcp | CodeGraphContext | |
|---|---|---|---|---|---|
| Semantic code search | Yes | Yes | Yes | Yes | Yes |
| Knowledge graph | Yes | Code-only | No | Code-only | Yes |
| Persistent AI memory | Yes | No | No | No | No |
| Cross-session recall | Yes | No | No | No | No |
| Symbol finding | Yes | Yes | No | Yes | Yes |
| Temporal time-travel queries | Yes | No | No | No | No |
| Conversation ingestion | Yes | No | No | No | No |
| Local-first / no cloud | Yes | Yes | Partial | Yes | Yes |
| MCP tools | 74 | 16 | ~5 | 14 | ~10 |
GitNexus (27k stars) has strong code search but no memory. claude-context (Zilliz, 5.9k stars) is Milvus-backed with no KG or memory. codebase-memory-mcp (DeusData) is the closest competitor — code + KG in a single binary — but has no AI memory layer.
Full Feature Matrix
| Feature | LocalNest | codebase-memory-mcp | GitNexus | claude-context | Basic Memory | Mem0 |
|---|---|---|---|---|---|---|
| Semantic code search (hybrid BM25+vec) | Yes | Yes | Yes | Yes | No | No |
| Knowledge graph (entities + triples) | Yes | Code-only | Code-only | No | No | No |
| Persistent AI memory | Yes | No | No | No | Yes | Yes |
| Symbol finding (defs/usages/callers) | Yes | Yes | Yes | No | No | No |
| AST-aware chunking | Yes | Yes | Yes | Yes | No | No |
| Temporal time-travel queries | Yes | No | No | No | No | No |
| Multi-hop graph traversal | Yes | No | No | No | No | No |
| Conversation ingestion | Yes | No | No | No | No | No |
| Agent-scoped isolation | Yes | No | No | No | No | No |
| Semantic dedup | Yes | No | No | No | No | No |
| Hooks system (pre/post callbacks) | Yes | No | No | No | No | No |
| Interactive TUI dashboard | Yes | No | No | No | No | No |
| Local-first / no cloud | Yes | Yes | Yes | Partial | Yes | Hybrid |
| MCP tools | 74 | 14 | 16 | ~5 | ~10 | 8 |
| Zero external deps | No (Node.js) | Yes (binary) | No | No | No | No |
LocalNest is the only server that checks every box in the first three rows.
Quick Start
# Install
npm install -g localnest-mcp
# Setup workspace + embeddings
localnest setup
# Verify
localnest doctor
Interactive dashboard:
localnest dashboard
MCP Client Config
After setup, add this to your AI client config:
{
"mcpServers": {
"localnest": {
"command": "localnest-mcp",
"startup_timeout_sec": 30,
"env": {
"MCP_MODE": "stdio",
"LOCALNEST_CONFIG": "~/.localnest/config/localnest.config.json",
"LOCALNEST_INDEX_BACKEND": "sqlite-vec",
"LOCALNEST_MEMORY_ENABLED": "true"
}
}
}
}
Works with Claude Code, Cursor, Windsurf, Cline, Continue, Gemini CLI, and any MCP-compatible client.
Tool Suites
LocalNest exposes 74 specialized MCP tools, organized into focused suites:
Workspace & Discovery — file navigation, project summaries, scoped reads
project_tree, read_file, file_changed, list_projects, list_roots, summarize_project
Search & Code Intelligence — hybrid search, symbols, AST-aware queries
search_hybrid, search_code, search_files, find, find_definition, find_usages, find_callers, find_implementations, get_symbol, rename_preview
Memory & Recall — persistent cross-session memory with semantic dedup
memory_store, memory_recall, memory_get, memory_update, memory_delete, memory_list, memory_store_batch, memory_delete_batch, memory_related, memory_suggest_relations, memory_add_relation, memory_remove_relation, memory_capture_event, memory_events, memory_status, memory_check_duplicate
Knowledge Graph — temporal triples, time-travel, multi-hop traversal
kg_add_entity, kg_add_triple, kg_query, kg_invalidate, kg_as_of, kg_timeline, kg_stats, kg_add_entities_batch, kg_add_triples_batch, kg_backfill_links, graph_traverse, graph_bridges
Organization — nests, branches, agent isolation, diary
nest_list, nest_branches, nest_tree, diary_write, diary_read
Ingestion & Hooks — conversation import, lifecycle callbacks
ingest_markdown, ingest_json, hooks_stats, hooks_list_events
Agent Context — priming, teaching, outcomes, task context
agent_prime, teach, capture_outcome, task_context, whats_new
System — health, indexing, updates, embedding status
health, server_status, index_project, index_status, embed_status, update_self, update_status, audit, backup, restore, help, usage_guide
Full parameter reference: Tool Documentation
Agentic Workflows
LocalNest is designed as the foundational context layer for AI coding agents:
- Cold start —
agent_primeinstantly hydrates the context window with relevant memories, recent changes, and project state. - Deep investigation —
findruns fused search across code fragments and historical design decisions in a single call. - Continuous learning —
teachsaves architectural rules that persist across sessions, ensuring agents never repeat mistakes. - Outcome capture —
capture_outcomerecords what worked and what didn't, building an experience base over time.
Enterprise-Grade Quality
- OIDC Trusted Publishing for verifiable npm provenance
- Continuous CodeQL static analysis on all branches
- OpenSSF Scorecard monitoring and proactive Dependabot updates
Troubleshooting
Installing from GitHub fails
Direct npm install -g git+https://... may fail with TAR_ENTRY_ERRORS. This is a known npm limitation.
Fix: clone, pack, install
git clone https://github.com/wmt-mobile/localnest.git
cd localnest && npm pack
npm install -g ./localnest-mcp-*.tgz
Semantic search not working
cd $(npm root -g)/localnest-mcp && npm install --no-save @huggingface/transformers
localnest doctor
Resources
- Documentation — Full tool reference, architecture, and guides
- Comparison — Detailed competitive analysis
- Architecture — Retrieval pipeline and memory graph internals
- Changelog — Release history
- Security — Vulnerability disclosure policy
Code intelligence. Knowledge graph. AI memory. One server. Your machine.
Как установить
Добавь это в claude_desktop_config.json и перезапусти Claude Desktop.
{
"mcpServers": {
"localnest-mcp": {
"command": "npx",
"args": []
}
}
}