SDL-MCP

Overview#

SDL-MCP (Symbol Delta Ledger MCP Server) is designed for AI coding agents to solve the token waste and low signal-to-noise ratio caused by reading entire files. Built on the MCP protocol, it acts as the perceptual layer for agents, supplying structured context rather than generating code directly.

Core Capabilities#

• Iris Gate Ladder: 4-rung progressive context control from ~100-token symbol cards to full source code, requiring agents to justify access escalation.
• Symbol Cards: Compact metadata (~100 tokens) for functions, classes, interfaces, types, and variables, with confidence-scored call resolution, community detection, and call chain tracing.
• Graph Slicing: Dependency graph traversal with weighted edges (call: 1.0, config: 0.8, import: 0.6) to return the most relevant N symbols within a token budget.
• Delta Packs & Blast Radius: Semantic diff beyond line-level git diff, with impact ranking, PR risk scoring, and amplifier symbol trend analysis.
• Live Indexing: Editor buffer changes pushed to in-memory overlay store, merged with persistent DB after background AST parsing, instantly reflecting unsaved changes.
• Agent Context: Task-shaped retrieval (debug/review/implement/explain) + description + token budget auto-selects the optimal Iris Gate path with feedback loops.
• Sandboxed Runtime Execution: 16 runtimes (Node.js/Python/Go/Java/Rust/Shell etc.) with executable whitelists, CWD restrictions, env var sanitization, timeouts, and smart output summarization.
• Development Memories (Opt-In): Graph DB-based memory system linking decisions/fixes to specific symbols; auto-invalidates stale memories on symbol changes; dual storage (graph DB + markdown files).
• SCIP Integration: Source Code Intelligence Protocol integration upgrading heuristic edges to compiler-verified precise edges (scip-typescript, scip-go, rust-analyzer, etc.).
• Tool Gateway: Projects 30 MCP operations into 4 namespaced tools (sdl.query, sdl.code, sdl.repo, sdl.agent).
• CLI Tool Access: Direct invocation of 30 operation aliases via sdl-mcp tool without an MCP server.

Use Cases#

AI-assisted understanding of large polyglot codebases, cross-file dependency debugging, PR semantic change assessment and risk scoring, code context extraction in CI/CD pipelines, team-level cross-session code decision memory.

Architecture#

• Transport: @modelcontextprotocol/sdk ^1.27.1
• Parsing: Tree-sitter (fork: @keqingmoe/tree-sitter) with 11 language grammar packages
• Graph storage: Kuzu (@ladybugdb/core 0.15.2, optional)
• High-performance: Rust native extensions via napi-rs (sdl-mcp-native, optional)
• Validation: Zod ^4.3.6; Observability: OpenTelemetry
• Optional inference: onnxruntime-node and tokenizers

Quick Start#

Requires Node.js ≥ 24.0.0:

npm install -g sdl-mcp --legacy-peer-deps
sdl-mcp init -y --auto-index
sdl-mcp serve --stdio

npx users can replace sdl-mcp with npx --yes sdl-mcp@latest.

Configuration#

• Point MCP clients to sdl-mcp serve --stdio; default Code Mode enabled (codeMode.exclusive: true)
• CLI: sdl-mcp tool <action> --<params> with JSON/stdin pipe and 4 output formats
• Memory: Opt-in via "memory": { "enabled": true }
• Runtime governance: custom whitelists, CWD limits, timeouts, env var sanitization
• SCIP ingestion: sdl.index.refresh

Notes#

Current version 0.10.8, authored by Gary Mullins. Licensed under SDL-MCP Community License v1.0 (free for personal/internal use; commercial distribution requires a paid license). Repository includes a VS Code extension directory (sdl-mcp-vscode/) and benchmarks, but marketplace links and third-party benchmark validation are unconfirmed.