dsh-mcp-workspace-scope

In DeepSeek Harness, MCP servers are registered process-globally, so every session drags the full tool list of every enabled server into its context window even when it only touches one. A Jira-only session still lugs three database drivers and a browser stack, burning tokens and risking accidental cross-calls. dsh-mcp-workspace-scope scopes MCP tool injection by the directory a session opens in: proj-a gets Atlassian, proj-b gets Playwright, everything else stays untouched. Subdirectories inherit their parent rule, longest-path wins, and matching happens at a separator boundary so /ws/proj never bleeds into /ws/project. It does not merely hide tool names; it hard-refuses the invocation at call time, turning a soft context saving into a real security boundary.

Beyond static rules, the plugin renders a persistent MCP badge in the composer tool row. Open it and you see which rule matched, each server live runtime state connected, starting, mount-failed, or disabled, and the visible versus hidden tool counts measured from that session own agent scope rather than predicted from the rules. Every server row carries a temporary toggle: need BigQuery for ten minutes? Flip it on, and the override lives only in that session memory, writes nothing to the rules file, and reverts the moment the agent ends. The badge turns amber when an allowed server cannot actually serve tools, and red when a rule names a server the profile no longer has, surfacing silent failures at a glance.

This plugin is a good fit for developers whose profile has accumulated several MCP servers but whose individual projects only need a subset. It subtracts only: it never starts, stops, or reconfigures a server, and it never touches secrets, command lines, or environment variables. Paired with the settings-page visual rule editor, where saving re-scopes running sessions immediately without a restart, it is a small, deterministic scalpel for managing MCP tool-injection granularity.