Developer Umut Sevdi has released Imza, a new terminal-based coding agent that aims to solve the 'black box' problem plaguing AI-assisted development. Unlike many competitors that bury users in endless streams of generated code, Imza enforces a three-mode workflow: Plan, Build, and Review. The tool is written in native C++, resulting in a lean ~20 MB binary that consumes between 5-8 MB of RAM at startup, a stark contrast to the resource-heavy Electron-based agents currently dominating the market.

Enforcing Developer Oversight

The core philosophy of Imza is keeping the human developer in the loop. In Plan mode, the agent investigates the repository and designs changes without modifying any files, treating the codebase as strictly read-only. This produces a Markdown specification document that remains alongside the conversation, allowing for iterative revisions and pinned notes. This ensures that the architectural decisions are vetted before a single line of code is written, preventing the 'summary drift' where developers lose track of implementation details.

Composable Tool Execution via Lua

Technically, Imza distinguishes itself by using sandboxed Lua scripts for tool execution rather than relying on separate model round trips for every read, search, or edit. This approach allows the model to compose bulk operationsβ€”such as filtering files and gathering contextβ€”into a single action, significantly reducing context overhead and latency. The agent supports up to eight concurrent subagents for research or build tasks, and includes a /undo command that reverts the last file change made by the agent, offering granular control over the mutation history.

Integration and Deployment Features

Imza supports bring-your-own-model configurations, including local OpenAI-compatible endpoints and the Anthropic Messages API. It features full Model Context Protocol (MCP) support and a 'sidechat' functionality for asking questions without disrupting the main agent's workflow. For automation, the tool includes a headless mode with --ask for read-only queries and --exec for file-modifying tasks, with scoped permission controls via --allow-dir and --allow-cmd flags. Installation is straightforward on macOS via Homebrew, with pre-built packages available for Linux and Windows, though macOS binaries currently require manual Gatekeeper approval.

Key Takeaways

  • Imza is a native C++ agent with a ~20 MB footprint, prioritizing low resource usage over Electron-based bloat.
  • The workflow enforces a read-only 'Plan' phase to generate Markdown specs before any code is modified.
  • Uses sandboxed Lua scripts to compose tool executions, reducing model round trips and context usage.
  • Supports headless CI/CD integration with strict directory and command permission flags.

The Bottom Line

Imza is a necessary correction to the 'generate and pray' mentality of current AI coding tools; by forcing a spec-first, read-only planning phase, it ensures developers actually understand the code they are shipping. While the 2-point Hacker News score suggests it is currently flying under the radar, the technical decision to use native C++ and Lua sandboxing shows a serious commitment to performance and developer control that the ecosystem desperately needs.