MCP Skill Hub: scoped context with visible evidence

A coding assistant has a finite context window. Large tool surfaces and unrelated project memories can use that space while making a request less clear. Retrieval is only useful if it selects relevant evidence and preserves the user’s intent.

The current design treats project identity and original source text as boundaries. A generated summary is useful context but cannot replace its source or become a new authorization.

The inspected context-service documentation describes an MCP entry point and an optional client prompt hook using the same deterministic retrieval service. The original prompt is retained separately from supplemental evidence.

Retrieval selects bounded project context and skill descriptions with source references and omissions. Full skill bodies remain available through explicit tools. SQLite supports the indexed context; optional indexing and curation sit outside the interactive prompt path.

When the service is unavailable or times out, the original prompt passes through. Missing project identity does not justify substituting context from another recently opened project.

The foreground context path needs no model call or API key. The client retains model choice, permission handling and continuation decisions. Automatic enrichment is opt-in; retired task interception and continuation behaviors are not part of the current path.

A retrospective alternative is broad automatic injection of all plausible context. It can increase recall while making irrelevant evidence and source confusion harder to detect. The documentation records an unpromoted broader selector experiment and requires whole-task evaluation before changing the foreground behavior.

The September 2026 measurement compares 87 versus 8 MCP tools and 16,362 versus 1,242 o200k_base tokens for compactly serialized schemas and server instructions: 92.41% less. This measures protocol payload, not total billed usage or development-task savings.

An isolated installed stdio smoke test checks integration, description retrieval, explicit content loading and prompt preservation. Synthetic marker tests assess source selection and scope boundaries, rather than general answer quality.

The published benchmark notes say whole-task promotion criteria have not passed. The case therefore reports the verified narrower result and leaves broader quality and savings claims open.

A smaller schema payload is a real measurement with a narrow meaning. It becomes a useful product outcome only if paired tasks in an installed client show preserved quality and practical benefit.

This work reinforced the distinction between evidence selection and agent authority. A source reference improves reviewability; it cannot authorize commands or manufacture a user instruction.