HIC Mouse

HIC Mouse is a VS Code extension and MCP-based toolkit built by HIC AI, Inc. that gives AI coding agents surgical precision when editing files. It addresses what the company calls "Execution Slop" — the failure mode where an AI agent makes the right plan and calls the right tool, but still produces a wrong or corrupted output. Mouse is available now across VS Code, Cursor, and Kiro, with a patent-pending approach to file editing.

What It Is

HIC Mouse is a precision file-editing layer that sits between an AI coding assistant and the filesystem. Rather than relying on the single string-replacement tool that most AI agents use by default, Mouse exposes a richer set of declarative editing operations delivered via the Model Context Protocol (MCP). It works with GitHub Copilot, Claude Code, Amazon Q Developer, Roo Code, Kilo Code, and other MCP-compatible assistants, and is model-agnostic across Claude (4.0+), Gemini (2.5+), GPT (4o+), Deepseek, Qwen3 Coder Next, and others.

Core Technical Approach

Mouse introduces three interlocking mechanisms that distinguish it from built-in agent file tools:

• Coordinate-based editing: Six declarative operations — INSERT, DELETE, REPLACE, REPLACE_REGION, FOR_LINES, and ADJUST — let agents address content by line number and character offset rather than echoing back large blocks of text. The company states this reduces output tokens by approximately 70%.
• Dialog Box staging with atomic rollback: Multi-operation and multi-file edits are staged in memory before touching disk. Agents can Save, Cancel, Inspect, or Refine staged changes. If any edit in a batch fails, all changes roll back automatically, leaving the codebase in a clean state.
• Tool-response engineering: Mouse embeds contextual guidance directly in every tool response — breadcrumb trails showing file structure (a topViewport field), next-action suggestions, blast-radius warnings, proactive staged-state reminders, and reason-field context threading. The company coined the term "tool-response engineering" to describe this embedded guidance layer.

Research-Backed Performance Claims

HIC AI publishes a technical report describing three preregistered confirmatory studies (67 paired trials) comparing Mouse against GitHub Copilot's default tools. According to the company's published results:

• Easy tasks (BX-504D, N=23): Mouse was 3.6× faster and 37% cheaper per task; p < 10⁻⁶.
• Medium tasks (BX-504B, N=25): 56% first-try correctness for Mouse vs. 0% for baseline; p = 1.22 × 10⁻⁴.
• Hard tasks (BX-701R, N=19): 89% task completion for Mouse vs. 0% for baseline; p = 7.63 × 10⁻⁶.

These figures are vendor-published claims from HIC AI's own research page and technical report.

Compatibility and Setup Path

Mouse installs from the VS Code Marketplace or Open VSX. After installation, users open the Command Palette, run "Mouse: Initialize Workspace", select their AI assistant, and reload the window — Mouse then auto-configures the MCP config file. System requirements are minimal: a supported IDE and Node.js 20+, with approximately 1.3 MB of disk space. The extension runs entirely locally; HIC AI states it does not collect or transmit source code, prompts, or AI responses.

Confirmed compatible IDEs include VS Code, Cursor, and Kiro. Windsurf, Visual Studio, JetBrains IDEs, and Google Antigravity are listed as pending. On the AI client side, GitHub Copilot, Amazon Q Developer, Claude Code, Roo Code, and Kilo Code are confirmed.

Roadmap and Current Direction

HIC AI's about page outlines three phases. The current phase covers the existing Mouse toolkit. The next phase includes advanced structured data support for CSV, JSON/YAML, and ASCII tables, expanded Business features (SCIM provisioning, SOC 2 compliance, audit logging, SSO enhancements), additional IDE and model support, and Mouse binaries in Rust for air-gapped and unattended scenarios without Node.js or VS Code. A future phase targets precision AI editing beyond code — covering Outlook, Microsoft Office documents, and Google Workspace equivalents. The GitHub repository was last updated in July 2026.