Projects / mcp-housekeeping-claude
Audit and safely clean the filesystem state Claude applications accumulate.
Codified audits across the three storage areas Claude apps leave on macOS: Claude Desktop and Cowork sessions, Claude Code state under the user's home, and VSCode chat sessions. Each audit step is its own tool, so an agent orchestrates the checks and writes the report. Every tool maps to an access level derived from its MCP annotations, and the default exposes read-only tools only.
Public mcp repository · source
The problem
What this is for.
Claude applications accumulate sessions, memory files, caches, artifacts and orphaned project state across three separate storage areas, and none of it is safe to delete blind — some of it is the only copy of a conversation you may still want. This server codifies the audits, so an agent can report what is actually there and prune only what is demonstrably obsolete.
Whose problem
Whether it is yours.
You want this if you use Claude Desktop, Cowork, Claude Code or the VSCode extension heavily enough that disk usage, session sprawl or stale memory files have become a real cost.
Capabilities
What it actually does.
- Claude Code: projects, sessions, memory files, storage summaries, obsolete-session detection, orphaned-project pruning and project relocation.
- Claude Desktop and Cowork: workspaces, sessions, artifacts health, plugin inventory, backups, reports and project cache status.
- VSCode chat sessions: workspaces, storage summary and obsolete-session detection.
- Session reading and checkpointing, so a conversation can be acquired before anything is pruned.
- Pruning tools for each area, separated from the audits that identify what is eligible.
State
How finished it is.
Public source, not published to any package registry, and macOS only — it addresses the paths those applications use there. Clone it, build it, and register the built server with your MCP client.
Using it
How to start.
Not yet published to a registry. Clone the repository, build it, and register the built server with your MCP client.
Boundaries
What it deliberately does not do.
Each audit step is its own tool, so the agent orchestrates the checks and writes the report rather than the server deciding what matters. Every tool maps to an access level derived from its MCP annotations, and the default exposes read-only tools only, so pruning requires a deliberate change. It observes and prunes runtime state; it decides nothing about knowledge.
Authority
What this repository owns.
A local hygiene adapter for agent-runtime state. It observes and prunes; it decides nothing about knowledge.
That boundary is deliberate. The ecosystem separates philosophy, engineering discipline, capability semantics, portable standards, and executable behaviour so that none of them can quietly absorb another. The full directory shows where this project sits relative to the rest.
Nearby
Other projects of the same kind.
mcp-ki-kb-fs— Give an agent scoped read and write access to local knowledge-base directories.mcp-git-audit— Audit a whole tree of Git repositories, read-only until an operator says otherwise.mcp-gsuite— Connect an agent to Google Workspace, with outbound mail confined to drafts.mcp-m365— Connect an agent to Outlook and OneDrive through the Microsoft Graph API.
See every public project, or go straight to the released command-line tools.