VSCode extension as Model Context Protocol (MCP) client that connects to MCP servers and exposes their tools as VSCode language model tools
This VSCode extension acts as an MCP client, not a server. It registers 5 tools via VSCode's languageModelTools API, not MCP. Tool definitions are present in package.json with schemas and descriptions, but quality is inconsistent. Three tools (tabCount, findFiles, runInTerminal) lack complete parameter descriptions. The two math tools (add, multiply) have minimal descriptions (18-19 chars, below the 34-char p10 baseline). No error handling guidance, no output schemas documented, and no tool annotations (readOnlyHint/destructiveHint). The runInTerminal tool is IRREVERSIBLE but has no confirmation pattern or dry-run support. Parameter descriptions are sparse or missing entirely.
Find files in the workspace matching a glob pattern
Run a command in a terminal and return the output
The number of active tabs in a tab group
Return result of adding two numbers, a + b
Return result of multiply two numbers, a * b
Math tools (add, multiply) have descriptions under 20 characters ('Return result of adding two numbers, a + b' = 44 chars is acceptable, but lacks context on WHEN to use them vs alternatives). Descriptions do not explain purpose or use case.
runInTerminal tool is IRREVERSIBLE (executes arbitrary commands) but has no confirmation pattern, dry-run support, or error recovery guidance. Agents can execute destructive commands without safeguards.
No output schemas documented for any tool. LLMs cannot predict return structure, forcing them to guess at response fields and plan downstream calls blindly.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
findFiles tool lacks pagination support (no limit, offset, or page parameters). A glob pattern matching thousands of files will return all results, potentially exhausting context window.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. Agents cannot distinguish safe reads from irreversible operations without explicit hints.