Stop feeding your AI agent raw source code. Give it structured intelligence instead.
🇧🇷 Ler em Português · 🇬🇧 English
Quick Start · Why MRCP · Tools · Live Demo · Contributing
Every time an AI coding agent reads your repository by opening file after file, it's paying what this project calls the "AI Tax": thousands of wasted tokens, slower responses, and a real chance of hallucinating about code it half-read.
MRCP Engine is a deterministic, AST-based intelligence layer that sits between your repository and your AI agent. Instead of dumping raw files into the context window, it parses your code once with Tree-sitter WebAssembly and hands back exactly the structured JSON your agent asked for — architecture graphs, security audits, type signatures, dead code, docs, and more.
flowchart LR
subgraph before["❌ Without MRCP"]
A1["AI Agent"] -->|"reads every file"| A2["3,000+ tokens/file"]
A2 --> A3["slow · expensive · guesses"]
end
subgraph after["✅ With MRCP Engine"]
B1["AI Agent"] -->|"one MCP / REST call"| B2["Tree-sitter AST parse"]
B2 --> B3["50–200 tokens · exact · fast"]
end
No LLM calls, no guessing — just deterministic parsing that turns an entire repository into a context package your agent can actually afford to read.
Auto-configure every MCP-compatible IDE and AI agent installed on your machine with a single command:
npx mrcp-engine setupThis detects and wires up the MCP server for:
| 🟢 Claude Desktop | 🟢 Claude Code | 🟢 Cursor | 🟢 Windsurf |
| 🟢 VS Code | 🟢 Antigravity IDE | 🟢 Gemini CLI | 🟢 OpenCode |
| 🟢 Ollama (MCP) | 🟢 Codex |
Prefer to wire it up by hand? Add this to your MCP client config:
{
"mcpServers": {
"mrcp-engine": {
"command": "npx",
"args": ["-y", "mrcp-engine@latest"]
}
}
}Or point any remote-HTTP-capable client (Antigravity, Cursor, Windsurf, VS Code) straight at the hosted instance — no install required:
{
"mcpServers": {
"mrcp-engine": {
"url": "https://mrcp-engine.vercel.app/api/mcp",
"transport": "streamable-http"
}
}
}No MCP support in your client? Every tool is also a plain GET/POST REST endpoint — see REST API reference.
| Raw context stuffing | MRCP Engine | |
|---|---|---|
| How it reads code | LLM re-reads full file text | Deterministic Tree-sitter AST parse |
| Token cost | ~3,000 tokens per average file | ~50–200 tokens per structured response |
| Consistency | Varies by prompt, can hallucinate structure | Same JSON schema, every time |
| Coverage | One file at a time | Whole-repo graph in one call |
| Non-code assets | Usually skipped | CSV/DOCX/XLSX/PDF/JSON/YAML/XML via mrcp_document_analyzer |
| Security/architecture checks | Manual prompting | Purpose-built audit tools |
MRCP doesn't replace your LLM's reasoning — it replaces the expensive, unreliable part of getting code into that reasoning in the first place.
- ⚡ Modular mode — call exactly the tool you need (
mrcp_document_analyzer,mrcp_security_compliance_audit,mrcp_code_metrics_health_scorer...) for a focused, single-purpose task. - 🚀 Full Suite mode — one call (
mrcp_run_full_repository_suite/GET /api/full-analysis) runs all 13 analysis tools in parallel, streams intermediate results tomrcp-analysis.json, and returns a consolidated executive report.
1. 🏗️ Core Engine & Document Intelligence — click to expand
| Tool | Endpoint | What it does |
|---|---|---|
analyze_repository |
GET /api/analyze?repo=<url> |
Structural AST graph: nodes (files/modules/functions), edges (dependencies), cyclomatic complexity, coupling, hotspots |
mrcp_document_analyzer |
GET /api/document-analyzer?repo=<url> |
Deterministic parser for non-code repos: CSV, TSV, TXT, MD, DOCX, XLSX, XLS, PDF (text layer), JSON, YAML, XML, LOG. Produces a knowledge graph, tabular schema inference with TypeScript interfaces, a Document Quality Index (0–100), and broken-link detection — no OCR |
get_repository_skills_contract |
GET /api/skills?repo=<url> |
Refactoring contracts for hotspot files (complexity > 50), enforcing a Zero Regression Policy on public signatures |
mrcp_run_full_repository_suite |
GET /api/full-suite?repo=<url> |
Runs all 13 tools in parallel and writes the executive dashboard |
2. ⚡ High-Efficiency Agent Offloading — click to expand
| Tool | Endpoint | What it does |
|---|---|---|
mrcp_api_contract_generator |
GET /api/api-contract?repo=<url> |
Extracts routes/handlers (Next.js, Express, Fastify, Hono, FastAPI, Flask) → OpenAPI 3.0.3 spec + typed TypeScript SDK, hallucination-free |
mrcp_monorepo_package_graph_analyzer |
GET /api/monorepo-graph?repo=<url> |
Maps pnpm/turborepo/lerna/nx topology, inter-package deps, topological build order, packages affected by a diff |
mrcp_docstring_api_doc_generator |
GET /api/doc-generator?repo=<url> |
Generates TSDoc/JSDoc/Python docstrings and Markdown API reference tables for undocumented public symbols |
mrcp_ast_refactor_applier |
POST /api/refactor-applier |
Batch AST refactors — rename symbols, extract interfaces, update imports across dozens of files — in ~30ms |
mrcp_type_signature_extractor |
GET /api/type-signature-extractor?repo=<url> |
Extracts only type signatures / .d.ts / Zod schemas, stripping function bodies (≈3,000 → 50 tokens) |
mrcp_git_diff_semantic_summarizer |
POST /api/diff-summarizer |
Strips formatting noise from diffs, summarizes changes by semantic domain at the AST level |
mrcp_dependency_compatibility_resolver |
GET /api/dependency-resolver?package=<name> |
Resolves SemVer compatibility, peer-dependency conflicts, and breaking-change risk against the real npm registry |
mrcp_dead_code_pruner |
GET /api/dead-code-pruner?repo=<url> |
AST reachability analysis for unused exports, dead variables, unused imports |
mrcp_sql_schema_orm_contract_generator |
GET /api/sql-orm-contract?repo=<url> |
Parses Prisma / SQL DDL / Drizzle / TypeORM into typed tables and queries |
3. 🛡️ Predictive & Security Engineering — click to expand
| Tool | Endpoint | What it does |
|---|---|---|
mrcp_code_metrics_health_scorer |
GET /api/code-health?repo=<url> |
Maintainability Index (0–100), technical-debt grade (A–F), cognitive-load distribution, top 5 refactor priorities with effort estimates |
mrcp_env_secret_contract_validator |
GET /api/env-validator?repo=<url> |
Maps process.env / os.environ usage, audits against .env.example, flags leak risk in public bundles, generates Zod validation schemas |
mrcp_impact_analysis |
POST /api/impact-analysis |
AST Blast Radius — downstream files/tests impacted by a change, before you commit |
mrcp_security_compliance_audit |
GET /api/security-audit?repo=<url> |
Static AST security audit: OWASP patterns, hardcoded secrets, unsafe shell execution, vulnerable deps, GPL license exposure |
mrcp_architectural_drift_detector |
GET /api/architecture-drift?repo=<url> |
Detects architectural drift, circular dependencies (Tarjan's algorithm), Clean Architecture violations |
mrcp_auto_test_coverage_gap_finder |
GET /api/test-gap-analysis?repo=<url> |
Maps high-complexity, untested functions and generates Vitest/Jest stubs |
mrcp_context_pruning_pack |
GET /api/context-pack?repo=<url>&task=<desc> |
Task-aware AST context slicing — 60–90% token savings |
4. 🌐 Web Search & Scraping — click to expand
| Tool | Endpoint | What it does |
|---|---|---|
mrcp_web_search |
GET /api/web-search?q=<query> |
Fast web search |
mrcp_web_scrape |
GET /api/scrape?url=<url> |
Clean text extraction, stripped of ads/nav/scripts |
mrcp_web_smart_search |
GET /api/smart-search?q=<query>&topN=2 |
Search + ranked scrape of the top N results |
mrcp_clone_page |
GET/POST /api/clone?url=<url> |
PageCloner Pro Engine: Reverse-engineers web pages into design tokens, semantic components, layout tree, purpose classification, completion needs, and an ultra-faithful AI prompt |
5. 👥 Triage & HR — click to expand
| Tool | What it does |
|---|---|
mrcp_triage_parse_resume |
Deterministic résumé parsing |
mrcp_triage_score_candidate |
Job-description-vs-candidate match score |
mrcp_triage_generate_hr_report |
Formatted technical opinion for HR |
Every MCP tool is also a plain HTTP endpoint — use it from any language, no MCP client required:
curl "https://mrcp-engine.vercel.app/api/analyze?repo=https://github.com/your-org/your-repo"Full endpoint table
| Endpoint | Method | Description |
|---|---|---|
/api/analyze?repo=<url> |
GET |
AST graph & complexity metrics |
/api/skills?repo=<url> |
GET |
Refactoring contracts for hotspots |
/api/api-contract?repo=<url> |
GET |
Route extraction, OpenAPI 3.0, TypeScript SDK |
/api/code-health?repo=<url> |
GET |
Maintainability Index & technical debt |
/api/env-validator?repo=<url> |
GET |
.env validation, leak alerts, Zod schema |
/api/monorepo-graph?repo=<url> |
GET |
Package dependency graph & build order |
/api/doc-generator?repo=<url> |
GET |
JSDoc/TSDoc generator & API reference |
/api/refactor-applier |
POST |
Batch AST refactors |
/api/type-signature-extractor?repo=<url> |
GET |
Type-signature-only extraction |
/api/diff-summarizer |
POST |
Semantic git-diff summary |
/api/dependency-resolver?package=<name> |
GET |
Version/compatibility resolution |
/api/dead-code-pruner?repo=<url> |
GET |
Dead-code detection |
/api/sql-orm-contract?repo=<url> |
GET |
Typed DB/ORM contracts |
/api/impact-analysis |
POST |
Change blast radius (body: { repoUrl, modifiedFiles }) |
/api/security-audit?repo=<url> |
GET |
Security & license audit |
/api/architecture-drift?repo=<url> |
GET |
Architecture drift detection |
/api/test-gap-analysis?repo=<url> |
GET |
Test-gap detection & stubs |
/api/context-pack?repo=<url>&task=<desc> |
GET |
Task-sliced LLM context package |
/api/clone?url=<url> |
GET/POST |
PageCloner Pro: Tokens, semantic components, layout & AI prompt |
/api/page-prompt?url=<url> |
GET |
Direct Markdown AI prompt for code reconstruction |
/api/full-analysis |
GET |
Full 13-tool diagnostic suite |
/api/mcp |
POST |
Central MCP endpoint (JSON-RPC 2.0) |
apps/vscode ships a native extension bringing MRCP's analysis straight into your editor:
- Sidebar — Quick Actions, Health Metrics (A–F / MI 0–100), Security &
.envAudit, Architecture & Dependencies, Test Gaps & Dead Code, Document Intelligence (DQI) - MRCP Cockpit — visual dashboard: complexity distribution, live KPIs, AST graph viewer, jump-to-code security/route tables
- Inline CodeLens — real-time cyclomatic complexity (
⚡ MRCP: Complexity 3 (Low 🟢)) plus a one-click Copy for AI button on every function/class/method - Problems panel integration — native warnings for exposed secrets and missing
.envvariables - AI Context Packer — one-click, ~95%-token-savings context packaging to paste into ChatGPT, Claude, or Cursor
pnpm --filter mrcp-vscode compile
cd apps/vscode && npm run package # → mrcp-vscode-2.6.0.vsixA hosted instance is already running — no setup needed to try it:
https://mrcp-engine.vercel.app
curl "https://mrcp-engine.vercel.app/api/code-health?repo=https://github.com/facebook/react"- 13-tool full diagnostic suite with parallel execution
- Native VS Code / Cursor / Windsurf extension
- Document intelligence for non-code repositories
- v3: on-demand file-content retrieval — request a specific file by the metadata MRCP already returned, and get its content back as structured JSON (including visual content where applicable)
- Expanded language grammars beyond the current 14+
Have a request? Open a feature issue.
Issues, feature requests, and PRs are welcome — see CONTRIBUTING.md for setup and guidelines, and CODE_OF_CONDUCT.md for community expectations.
If MRCP Engine saves your agent (or your wallet) some tokens, a ⭐ star genuinely helps other people find it.
MIT © faelscarpato