AI Agent Monitor & Token Optimizer for Claude Code, Copilot, and Cursor. Monitors token counts, costs, cache efficiency, and provides prompt compression techniques. Includes pre-tool-use hooks for multiple AI agent platforms to optimize file reads and search operations.
Terse is not an MCP server. It is a Tauri desktop application (src-tauri/Cargo.toml) with a VSCode extension (vscode-extension/package.json) and Node.js backend (package.json). The 13 'tools' listed are shell script hooks and optimization utilities for third-party AI agents (Cline, Copilot, Cursor, Windsurf), not MCP tool definitions. No MCP protocol implementation is present. Tool definitions lack formal JSON Schema registration, are scattered across shell scripts, and have no structured input/output contracts. This is a prompt optimization framework, not an MCP server.
File globbing utility. Generally efficient, passes through without modification.
Cap results and context lines to prevent excessive token consumption. Limits head_limit based on output mode (200 for content, 50 for files_with_matches) and caps context lines to 5.
Cap line limit to prevent huge file reads. Default limit is 2000 lines. Skips binary/generated files and caps large files to 500 lines.
GitHub Copilot CLI preToolUse hook for grep/search. Caps search results to 100 to prevent excessive token consumption.
Cursor preToolUse hook for grep_search. Caps search results to 100 to prevent excessive token consumption.
Cline PreToolUse hook for list_files. Warns on broad recursive directory listings to prevent excessive token consumption.
Not an MCP server. Terse is a Tauri desktop app + VSCode extension that optimizes prompts for third-party agents (Cline, Copilot, Cursor, Windsurf). No MCP protocol implementation, no tool registration via MCP, no remote accessibility.
Tool definitions are shell script hooks, not formal MCP tool schemas. No JSON Schema input/output contracts, no structured parameter validation, no error classification.
Duplicate tool names across different agent platforms (read_file appears 4 times, grep appears 2 times, view appears 2 times). In an MCP context, this would violate tool uniqueness and confuse LLM selection.
Glob tool has empty input schema and 20-char description ('File globbing utility. Generally efficient, passes through without modification.'). Violates minimum description length (10 - 1024 chars) and provides no actionable guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
Cursor preToolUse hook for read_file. Caps input parameters to prevent excessive token consumption. Blocks binary/generated files and caps large files.
Cline PreToolUse hook for read_file. Blocks reads of very large/generated files and injects guidance context to prevent excessive token consumption.
Codex CLI PreToolUse hook for file read. Caps input parameters to prevent excessive token consumption. Blocks binary/generated files and caps large files.
Windsurf pre_tool_use hook for file reads. Blocks reads of generated/binary files that waste tokens.
Codex CLI PreToolUse hook for search/grep. Caps search results to 100 to prevent excessive token consumption.
GitHub Copilot CLI preToolUse hook for file view. Caps lines to prevent excessive token consumption. Blocks binary/generated files.
Windsurf pre_tool_use hook for view_file. Blocks reads of generated/binary files and extremely large files.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract structured data. Violates pattern:tool and pattern:response-shaper.