Most AI coding harnesses add dependencies for every feature they support. MaskShift takes the opposite approach: it runs on Node.js 22 using only built-in modules, including the interface renderer, and ships with 149 native tools and 44 bundled skills without requiring a single npm install.

What zero dependencies means in practice

The claim of zero runtime dependencies is precise. MaskShift does not run npm install because there is nothing to install. The interface renderer is built on Node's built-in modules. The tool implementations use Node's standard library. There is no HTTP server, no browser, no listening socket. The installer copies the code to a local directory and links a binary, and the tool runs.

For developers who have watched their node_modules directories grow to thousands of packages, the appeal is obvious. No dependency tree to audit. No supply chain risk from packages you did not choose. No version conflicts between tools. MaskShift is a single codebase that runs with whatever Node.js version you have installed.

The tool supports any instruction-following model through Ollama or other configured providers. It detects when a model lacks a native tool API and switches to an in-prompt text protocol automatically, so the full harness works regardless of which model you point it at.

149 tools, lazy loaded

The tool catalog covers filesystem operations, shell and process control, search and indexing, Git worktrees and checkpoints, language server protocol, browsers over Chrome DevTools Protocol, containers and Kubernetes, SSH and rsync, databases, runtimes, images, PDF and Jupyter notebooks, web retrieval, plugins, automations, memory, and orchestration. The catalog is always available to the harness, but only the capabilities relevant to the current step get inserted into the model's context window.

This lazy loading matters for token efficiency. A harness that dumps all 149 tool descriptions into every request wastes context on tools the model will not use for the current task. MaskShift evaluates what the task requires and loads only those tools, keeping the context window focused on the work at hand.

The 44 bundled skills are loaded lazily by description, alongside skills imported from Claude, Codex, Copilot, and workspace skill directories. Skills are higher-level workflows that combine multiple tools into coherent tasks, and the lazy loading ensures they do not clutter the context until the model needs them.

MCP fabric that connects on demand

MaskShift includes a lazy MCP fabric that supports both stdio and Streamable HTTP transports, stateless and legacy initialization, resources, prompts, qualified tools, imported configs, and the live official MCP Registry. Servers connect on demand, so the catalog never floods the context window.

This is a different approach from MCP implementations that connect to all configured servers at startup and keep them in memory. MaskShift connects when a tool from that server is needed, uses it, and disconnects. The result is lower memory usage and a context window that reflects only what the current task requires.

The integration with the official MCP Registry means you can browse and install servers without leaving the harness. The registry pull happens when you request it, not at startup, keeping the initial load fast.

Autonomous repo context

MaskShift builds an autonomous understanding of the repository you are working in. It reads project instructions, parses manifests, indexes the repository tree, builds indexed code chunks using lexical and optional semantic retrieval, maintains stored memory, tracks recent history, and monitors Git state. This context accumulates over time and persists across sessions.

The memory system is cost-aware. It uses Anthropic prompt-cache breakpoints on the stable prefix, decay and access-aware ranking for memory items, and a usage report tool that prices spend from a user-editable table rather than guessing at token costs. You know exactly what each interaction costs, and the memory system prioritizes items that are both relevant and cheap to include.

For teams working on large repositories, the indexed code chunks provide fast retrieval without loading entire files into context. The lexical plus semantic retrieval covers both exact matches and conceptual similarities, giving the model a complete picture of the relevant code without the overhead of reading every file.

Parallel agents and scheduled work

MaskShift supports independent sessions and optional isolated Git worktrees for delegated work. A main agent can spawn parallel agents that work on separate tasks, each with its own context and Git state. The worktrees prevent conflicts between agents editing the same files.

Scheduled work runs agent calls, direct tool calls, or host shell commands on an interval, a cron expression, or a one-shot timestamp. The daemon mode runs as a resident scheduler without the interface, suitable for systemd services or Docker containers. Automations can be created through the interface or the command line, and every automation is inspectable and editable.

The interface

The full-screen terminal interface has six views: transcript and composer, workspace file tree, searchable tool and skill catalog, MCP server network, automations and plugins, and the host shell. Every action is accessible through a fuzzy command palette, and every subcommand works headlessly with --json output.

The interface adapts to terminal size. Below 108 columns, the side rail hides and the header sheds telemetry. Below 80 columns, it degrades cleanly. Mouse support is optional and configurable. Color can be disabled entirely for environments that need it.

The design principle is that everything visible in the interface is also accessible from the command line. Nothing is hidden behind keyboard shortcuts that only work in the interactive mode. The command line is the interface, and the interface is the command line.

For developers who want a coding harness that does not add dependencies, does not flood the context window, and does not hide functionality behind configuration files, MaskShift provides a concrete alternative. It is MIT licensed and available on GitHub.