Mouse Brings Surgical Precision to AI Coding Agents

The Announcement

HIC released Mouse, a file-editing tool designed specifically for AI coding agents. The project was posted on Hacker News with a link to hic-ai.com. It replaces the single string-replacement approach common in current agents with coordinate-based operations, staged edits, and structured tool responses. The announcement includes a 14-day trial and documentation for the patent-pending system.

Coordinate-Based Editing Operations

Current agents typically receive one primitive: replace a string match with new text. This breaks when context shifts or multiple similar strings exist in a file. Mouse instead exposes six declarative operations that accept line and column coordinates.

INSERT places new content at an exact position. DELETE removes a defined range. ADJUST modifies whitespace or indentation without touching logic. Three additional operations handle block moves and comment toggling. Each call requires explicit start and end coordinates rather than pattern matching.

Agents must therefore track file state more carefully before issuing commands. The syntax stays close to natural language descriptions while remaining machine-readable. In practice this reduces off-by-one errors that occur when an agent guesses string boundaries from surrounding code.

Staged Changes with Atomic Rollback

Mouse does not write directly to disk on the first tool call. Edits enter a staging area that the agent can inspect, refine, or discard. Four control actions are available: Save commits the staged diff, Cancel discards it, Inspect returns the diff plus surrounding context, and Refine lets the agent issue additional coordinate operations against the same staging buffer.

Atomic rollback is guaranteed because the original file bytes remain untouched until Save. If an agent produces a broken intermediate state, it can revert without manual git checkout. This workflow mirrors a review step that human developers perform but is executed inside the agent loop.

Embedded Guidance in Tool Responses

Every Mouse response includes four fixed sections: contextual guidance, suggested next actions, risk assessment, and a viewport summary of the file tree. The guidance text points out potential side effects such as broken imports or scope changes. Risk assessment assigns a numeric score based on edit size and proximity to control flow statements.

Agents that lack persistent memory benefit from the viewport summary, which lists nearby functions and imports without requiring a separate read call. The added tokens increase response size but reduce follow-up tool invocations that would otherwise be needed to re-establish context.

Integration Considerations

Mouse runs as a local server that agents connect to through a defined JSON protocol. Existing frameworks that already support custom tools can register the six operations and four control actions. No changes to the underlying language model are required.

Performance overhead comes mainly from the extra tokens in each response and the staging buffer management. For agents that already perform many small edits, the coordinate model can lower total round trips once the agent learns to plan larger, precise operations.

Developers using Mouse must expose file paths and line numbers to the agent, which raises the same permission questions that apply to any tool with write access.

Limitations

The coordinate system assumes the agent maintains accurate line and column state. If an earlier edit shifts line numbers and the agent does not recalculate, subsequent calls target the wrong locations. The staging mechanism mitigates some damage but does not eliminate the need for correct bookkeeping.

Documentation does not yet specify conflict handling when multiple agents edit the same file concurrently.

FAQs

Does Mouse require changes to the language model itself? No. It registers as additional tools that any agent framework supporting custom function calls can load.

How does atomic rollback differ from git revert? Rollback happens inside the staging buffer before any commit occurs, so no git history is created for discarded attempts.

Is the coordinate syntax tied to a specific programming language? The operations work on raw text offsets and apply to any file type, though agents still need language awareness to choose correct coordinates.

---

📖 Related articles

• Agentic Coding: Una Trappola per lo Sviluppo Software?
• File agents.md: utili per gli agenti di coding?
• Phantom su GitHub: L'AI co-worker auto-evolvente e sicuro