LoomMCP — Cut Claude Bills by 95%. Keep Your Edge.

"We're building the foundation for what context compilation looks like in 2025 and beyond." — This repo

LoomMCP is a universal Model Context Protocol server that gives coding agents surgical access to your codebase. Instead of dumping 50k tokens of boilerplate on every request, it serves AST skeletons on-demand — achieving 95% token reduction while keeping code quality intact.

tl;dr

$20/day Claude bills  →  $1/day  (with 95% token reduction)
50k tokens/turn        →  2.7k tokens/turn  (same capability)

Install:

npm install -g @loom-mcp/server
loom start
/mcp add loom localhost:8080

The Problem

Every autonomous agent — Claude Code, Codex, Cursor Agent, Gemini Code Assist — suffers from the same curse: static context bloat. On turn 1, Claude reads your node_modules. On turn 50, it reads your node_modules again. You're paying $15/1000k tokens to re-read files that never change.

What LoomMCP Does Differently

• On-demand AST: Only the symbols the agent needs are paged in
• Session awareness: Tracks what's been focused this session
• Cross-session memory: Remembers code insights across sessions
• Task-aware compression: Debug mode keeps stack traces. Explore mode is minimal.
• Live dashboard: Real-time token savings at localhost:2337
• Session replay: JSONL playback for debugging agent behavior

Token Reduction Benchmark

node eval/benchmark.js .

| Repository | Files | Raw Tokens | TOON Tokens | Reduction | Latency | |:-----------|------:|----------:|------------:|:---------:|:-------:| | loommcp (self) | 39 | 55,813 | 2,729 | 95% | ~800ms |

Quickstart

# Install
npm install -g @loom-mcp/server

# Start the server
loom start

# Connect your client (see SETUP.md for all clients)

Claude Code

/mcp add loom localhost:8080

Claude Desktop (macOS/Windows)

Add to config:

{
  "mcpServers": {
    "loom": {
      "command": "npx",
      "args": ["-y", "@loom-mcp/server", "start"]
    }
  }
}

Cursor, VS Code, Codex, Zed, Windsurf, Gemini

See SETUP.md for all client configurations.

MCP Tools (28+ Total)

| Tool | Description | |:-----|:------------| | loom_get_topology | AST skeleton of entire codebase (~3k tokens) | | loom_focus | Page in a specific file or symbol body | | loom_search_symbols | Search symbols with relevance ranking | | loom_get_symbol | O(1) byte-offset symbol retrieval | | loom_find_importers | Reverse dependency graph | | loom_blast_radius | Change impact analysis | | loom_search_refs | AST-aware find references | | loom_hybrid_search | Keyword + semantic RRF fusion | | loom_remember | Store insight across sessions | | loom_recall | Retrieve cross-session memory | | loom_session_compress | Task-aware compression (debug/feature/explore) | | loom_diff_compress | Compressed git diffs | | loom_get_active_diff | Session changes since start | | loom_blur | Remove file from focus | | loom_get_symbol_importance | Symbol centrality scoring | | loom_get_changed_symbols | Files changed vs last commit | | loom_get_untested_symbols | Heuristic test coverage detection | | loom_get_deps | Dependency graph (text/DOT/JSON) | | loom_get_metrics | Session metrics + tool breakdown | | loom_get_sessions | Historical session data | | loom_bm25_search | BM25 ranking (statistically optimal search) | | loom_fuzzy_search | Fuzzy matching with Levenshtein scoring | | loom_find_dead_code | Detect unused functions and unreachable code | | loom_get_class_hierarchy | Class inheritance traversal | | loom_pagerank_centrality | Architectural importance via PageRank | | loom_attach_repo | Attach additional repos for cross-repo search | | loom_workspace_search | Search across all attached repos | | loom_semantic_search | Semantic embedding search (GPU optional) | | loom_detect_frameworks | Detect Nuxt/Next.js/React/Django/etc. | | loom_get_confidence | Methodology disclosure (confidence levels) | | loom_enforce_hook | Register PreToolUse/PostToolUse hooks | | loom_agent_info | Get supported agent integrations |

Live Dashboard

Starts automatically at localhost:2337 when you run loom start:

• /badge.svg — Embeddable reduction badge for your README
• /replay — Animated JSONL session playback
• /health — Readiness probe for production deployments
• /api/summary — JSON stats for dashboards/CI

LoomPacks

Pre-configured toolkits for frameworks. Install with:

loom pack list         # Show available packs
loom pack install react   # Install React pack
loom pack install node     # Install Node.js pack

Available: React, Node.js, Python Django, Rust, Go, Java Spring, C# .NET

System Prompts

Specialized prompts for different tasks:

• prompts/agent.txt — General coding assistant
• prompts/reviewer.txt — Code review workflow
• prompts/debugger.txt — Debug-first approach

Architecture

┌──────────────┐         ~2.7k tokens TOON         ┌───────────────┐
│   Claude     │ ◄─────────────────────────────────►│   LoomMCP     │
│   (Agent)    │                                   │  (AST Engine) │
└──────────────┘                                   └───────────────┘
                                                             │
                                                             ▼
                                                    ~56k raw tokens
                                                             ▼
                                                   ┌───────────────┘
                                                   │  Filesystem   │
                                                   └───────────────┘

Supported Languages

| Language | Status | |:---------|:-------| | TypeScript / JavaScript | ✅ Primary | | Python | ✅ Full | | Rust | ✅ Full | | Go | ✅ Full | | Java | ✅ Full | | C# | ✅ Full |

Competitive Analysis

| | LoomMCP | jCodeMunch | Srclight | |:--------|:-------:|:----------:|:--------:| | Token Reduction | 95% | 95-99.6% | ~90% | | Tools | 28+ | 11 | 42 | | Languages | 7 | 6 | 25+ | | Cross-Session Memory | ✅ | ❌ | ❌ | | Session Replay | ✅ | ❌ | ❌ | | Live Dashboard | ✅ | ❌ | ❌ | | Task-Aware Compression | ✅ | ❌ | ❌ | | Byte-Offset Retrieval | ✅ | ✅ | ✅ | | BM25 Search | ✅ | ✅ | ✅ | | Fuzzy Search | ✅ | ✅ | ❌ | | Dead Code Detection | ✅ | ✅ | ❌ | | PageRank Centrality | ✅ | ✅ | ❌ | | Class Hierarchy | ✅ | ✅ | ❌ | | Multi-Repo Workspaces | ✅ | ❌ | ✅ | | GPU Embeddings | ✅ | ❌ | ❌ | | Framework Providers | ✅ | ✅ | ❌ | | Enforcement Hooks | ✅ | ✅ | ❌ | | Methodology Disclosure | ✅ | ✅ | ❌ | | Agent Support (NOUS Hermes) | ✅ | ❌ | ❌ | | MIT License | ✅ | ❌ | ❌ | | LoomPacks | ✅ | ❌ | ❌ |

Where LoomMCP Leads

1. Most Complete Feature Set — 28+ tools beating jCodeMunch (11) and matching Srclight (42)
2. Cross-Session Memory — No competitor stores code insights between sessions
3. Session Replay — JSONL-based animated playback for debugging agent behavior
4. Live Dashboard — Real-time SSE visualization at localhost:2337
5. Task-Aware Compression — Adapts output based on task type
6. LoomPacks — Pre-configured frameworks, one command install
7. Enforcement Hooks — PreToolUse/PostToolUse hooks to force agent tool usage
8. NOUS Hermes Support — Native MCP client + server for Hermes agent
9. Methodology Disclosure — Confidence levels in every response
10. GPU Embeddings — Optional semantic search with GPU acceleration
11. Class Hierarchy — Inheritance tree traversal
12. PageRank Centrality — Architectural importance scoring

Production Features

• Circuit breaker: Won't hammer your filesystem
• Path traversal protection: Can't be tricked into reading ../../.env
• File watchers: Auto-invalidate cache when files change
• Session recorder: JSONL persistence in .loom/sessions/
• Config file: loom.config.json for workspace customization
• Docker: Multi-stage build with healthcheck

Docker

docker build -t loom-mcp .
docker run -p 2337:2337 -v $(pwd):/workspace loom-mcp

CI/CD

Every push to main auto-builds, tests, and publishes to npm.

Requirements

• Node.js 18+
• Works on macOS, Linux, Windows

Roadmap

• [x] Phase 1: Foundation — 17 tools, 7 languages, 95% reduction, dashboard
• [x] Phase 2: Expansion — Dependency graph, session metrics, client adapters
• [x] Phase 3: Scale — LoomPacks, system prompts, CI/CD, Docker, 32 tests
• [x] Phase 4: Dominate — 28+ tools, BM25, fuzzy, dead code, PageRank, multi-repo, GPU, enforcement hooks

License

MIT — use it forever, no fees, no vendor lock-in.

Bottom line: LoomMCP matches jCodeMunch's 95% token reduction while adding genuinely novel features — all MIT licensed, all local, all fast.