LoomMCP is the context compiler for AI coding agents. It indexes your codebase once and delivers exact symbol retrieval instead of dumping entire files.
Works with Claude Code, Cursor, VS Code, Codex, OpenCode, and any MCP client.
Reads entire files. Wastes tokens.
Gets exact symbols. Saves 97%.
We parse and index your codebase into a symbol database. Fast. Local. Private.
Agents search exact symbols, references, and relationships instead of reading entire files.
Cross-session memory remembers what was learned. No repetitive context. Massive savings.
Remembers context across sessions and projects.
Different modes for debug, explore, build, and review.
Real-time token savings and usage analytics.
Replay and debug agent sessions step-by-step.
Semantic + lexical search for best results.
Visualize relationships between files and symbols.
Understand impact before you change anything.
Logging, health checks, configurable, and secure.
LoomMCP ships with a full observability dashboard at localhost:2337. Track token savings, focused files, tool call history, and live events — all persisted across sessions.
Real baseline vs compressed token counts. Persisted across sessions.
Which files Claude is currently focused on, with line and token counts.
Real-time SSE stream of every MCP tool call as it happens.
Browse past sessions, tool call breakdowns, and token totals.
The Topology page shows the last AST skeleton Claude fetched. Every function signature, class, and type — compressed to TOON format. 97% fewer tokens than raw files.
The Active Lens tracks every file Claude has focused into context. See exact token counts per file, dependency depth, and how much of your 20-file budget is used.
Set your workspace root, focus budget, max topology depth, and auto-refresh behaviour. All settings persist and take effect immediately — no restart required.
Here's exactly what happens in the dashboard when Claude uses LoomMCP to diagnose and fix a login bug in a 40,000-token TypeScript codebase.
LoomMCP scans all TypeScript files, generates TOON skeletons (function signatures only), and returns them in ~16ms. Dashboard records: raw 54,932 tokens → 1,456 TOON tokens.
token_reduction: 97% · latency_ms: 16
Only the exact function body is paged in — 42 lines, 1,204 tokens. The other 2,038 lines of auth.ts stay out of context.
focused: src/auth.ts::loginUser · lines: 42 · token_estimate: 1204
AST-aware reference search finds 14 call sites across 8 files in milliseconds — no grep, no file reads.
symbol: loginUser · refs: 14 · scope: workspace
The diff is compressed and scoped to only changed symbols. The dashboard Diff page shows the exact additions and deletions without noise.
files_changed: 1 · additions: 3 · deletions: 1
Includes topology, symbol search, symbol fetch, importers, blast radius, references, active diff, blur, hybrid search, memory, compression, diff compression, dependencies, metrics, and sessions.
Measured using the same methodology as jCodeMunch — tiktoken cl100k_base encoding.
| Repository | Files | Raw Tokens | TOON Tokens | Reduction |
|---|---|---|---|---|
| loommcp (self) | 33 | 54,932 | 1,456 | 97% |
| eval/fixtures/large_oss | 4 | 6,413 | 108 | 98% |
| eval/fixtures/medium_webapp | 3 | 3,318 | 69 | 98% |
| eval/fixtures/small_api | 3 | 1,013 | 23 | 98% |
Average token reduction: 97% (tiktoken cl100k_base)
Run: npm run build && node eval/benchmark.js . --json
"LoomMCP cut our Claude bill by 97%."
"The cross-session memory is incredible."
"Finally, an MCP server that just works."
Join thousands of developers saving millions of tokens every day.