tracedecay

TraceDecay runs as a local daemon-managed service that eliminates the need for AI coding agents to repeatedly run grep, glob, and raw file read operations when working with a codebase. It builds a structured semantic graph of the target repository, storing indexed data via a daemon-owned SQLite runtime and embedded graph store, with local operation as the default behavior. The tool exposes a suite of typed Model Context Protocol (MCP) tools grouped across common coding workflows: discovery tools for context retrieval, symbol search, and source outline lookup; graph traversal tools to map callers, callees, impact radius, and affected files; code health tools to measure complexity, detect dead code, identify unmounted files, assess coupling, and flag test risk; git workflow tools to pull diff context, PR context, changelogs, and test mappings; anchored single-file edit tools that trigger automatic re-indexing after changes; and persistent project memory tools to store, search, and manage project facts across branches and linked git worktrees. It ships native integrations for a range of popular AI coding hosts including Claude Code, Codex, Cursor, Cline-family tools, Kiro, Kimi Code, OpenCode, and Hermes, with both global and project-local installation options. The background daemon manages index freshness, background convergence, and immutable generation provenance across worktrees, branches, and commits, reporting explicit state for warming, refresh-required, partial, or unavailable index coverage rather than running hidden syncs or silently using stale index data. Supporting features include a local web dashboard for graph exploration, project memory review, LCM session search, token savings tracking, and cost analytics, plus a terminal UI for live monitoring of MCP call savings and costs. Installation is supported across macOS, Linux, and Windows, with prebuilt release archives, an install script that verifies build provenance via GitHub attestation, and built-in self-upgrade commands. Privacy controls are built into the design: core indexing, query, memory, and dashboard functions run locally by default, with optional network features (aggregate token savings counter, release checks, semantic model artifact downloads from Hugging Face, remote authority replication) that require explicit opt-in and fail closed if credentials or policy checks fail. The project is implemented in Rust, with tree-sitter based language extractors available in configurable feature tiers, and is released under the MIT license.