SCS Brings Bounded Repository Intelligence to Coding Agents via MCP

Summary

SCS is a headless code-intelligence service for coding agents. It enrolls a repository only after an explicit CLI, MCP, or client request, builds a structural and semantic index, and exposes repository investigation through the query_code MCP tool. The service starts with an empty index and does not modify repository source; Git-visible changes are watched and reindexed in the background after enrollment. Queries select one of seven bounded playbooks, optionally with the local Laya classifier, and return evidence together with routing, trace, and completeness information. Fast, balanced, and thorough modes impose fixed time and evidence budgets. SCS supports Apple Silicon macOS and x86-64 Linux with CPython 3.14, and indexing requires an embedding provider. The default configuration uses OpenAI's text-embedding-3-large API with 3,072 dimensions and sends source-derived entity text, while local OpenAI-compatible servers and an in-process MLX provider are also supported. SCS says it does not send whole repository files to a summarization service. Laya is an optional Apple Silicon MLX worker that receives the goal and explicit anchors, not repository source, embeddings, or retrieved evidence; failed inference falls back to deterministic routing. The release installer verifies its wheel, constraints, and required uv binary, installs without sudo, and requires users to verify a checksum manifest, although current macOS releases are not Apple-signed or notarized. Operational commands cover project enrollment, reingestion, deletion of SCS-owned derived state, health checks, metrics, and daemon status. MCP bridges share a lazily started daemon, while persistent state is stored under SCS_HOME.