The full toolbox for any coding model, in one zero-dependency terminal harness.
167 tools, 50 skills, persistent code-graph intelligence, a live clickable browser view and
worktree-isolated multi-agent orchestration — but only what the current step needs ever
touches the model's context.
MaskShift gives a coding model one control plane for repository understanding, file edits, the host shell, Git recovery, language servers, browsers, containers, databases, remote machines, memory, scheduled work, plugins, external coding agents, skills and MCP servers. The full catalog is always available to the harness; only the capabilities relevant to the current step are inserted into model context.
It runs on Node.js 22 using only built-in modules — the interface renderer included. No npm runtime dependency tree, no HTTP server, no browser, no listening socket.
Most terminal coding agents pick a model vendor and a fixed tool list. MaskShift is built the other way around: the model is a config line, and the tool surface is the whole catalog, held out of context until a step actually needs it.
| MaskShift | Claude Code | OpenCode | Codex CLI | Nous Hermes Agent | |
|---|---|---|---|---|---|
| Model support | Any provider — Ollama, OpenAI, Anthropic, Gemini, OpenRouter, LM Studio, vLLM — plus a text protocol for models with no tool API | Anthropic's Claude family | Any provider, 75+ models | OpenAI's GPT family, via ChatGPT plan or API key | Any provider — Nous Portal, OpenRouter, OpenAI, custom endpoints |
| License | GPL-3.0, open source | Proprietary (Commercial Terms of Service) | MIT, open source | Apache-2.0, open source | MIT, open source |
| Runtime dependencies | 0 — Node.js built-ins only | npm dependency tree | Go binary + provider SDKs | Rust binary, containerized sandbox by default | Python 3.11 + Node.js, ripgrep, ffmpeg |
| Native tool surface | 167 tools — shell, LSP, browsers, containers, Kubernetes, SSH, databases, PDFs, images, MCP | Built-in file/shell/web tools, plus MCP | File/shell tools plus per-language LSP and MCP | File/shell/sandboxed-exec tools, plus MCP | 40+ tools, plus MCP over the agentskills.io standard |
| Codebase intelligence | Persistent file/symbol/call graph with reverse change-impact analysis | Ad hoc search and grep | Per-language LSP (go-to-definition, references) | Ad hoc search and grep | General-purpose tool search, not code-graph specific |
| Multi-agent orchestration | Executable dependency DAGs, worktree-isolated by default, bounded parallel waves | Subagents (Explore, Plan, general-purpose) | Multiple parallel sessions | Native subagents, sandboxed | Isolated subagents, each with its own conversation and terminal |
| Interface | Full-screen, zero-dependency TUI — eight views including a live, clickable browser tab and a dedicated Git view, mouse and keyboard — plus an MCP server so any MCP client can drive it | Chat-style terminal interface, plus IDE and desktop apps | Terminal-first TUI, plus desktop and IDE extensions | Terminal chat, plus a VS Code extension | Full TUI, plus Telegram, Discord, Slack and other surfaces |
| Primary focus | A coding harness — any model, the whole tool catalog | Coding agent | Coding agent | Coding agent | General-purpose autonomous assistant; coding is one surface among many |
Not a knock on any of them — they make different tradeoffs on purpose. This is where MaskShift lands.
Features · Install · Run it with Ollama · Configure · The interface · The command line · Documentation · Contributing · Development · Deployment
| 167 native tools | Filesystem, shell and process control, search and indexing, Git worktrees and checkpoints, LSP, browsers over CDP, containers and Kubernetes, SSH and rsync, databases, runtimes, images, PDF and Jupyter, web retrieval, plugins, automations, memory and orchestration. → tool inventory |
| 50 bundled skills | Loaded lazily by description, including 14 Apache-2.0 skills imported from Anthropic, alongside skills from Claude, Codex, Copilot and workspace directories. → skills |
| Lazy MCP fabric | stdio and Streamable HTTP, 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. MaskShift is also an MCP server itself — maskshift mcp serve exposes its native tool catalog over stdio to Claude Desktop, Claude Code, an IDE, or another MaskShift instance. → MCP config |
| Any model | Ollama, OpenAI Responses, OpenAI-compatible servers, Anthropic, Gemini, OpenRouter, LM Studio and vLLM — with a text protocol that gives models without a native tool API the full tool surface. → providers |
| Codebase intelligence | Persistent file/symbol/import/call graph, reverse change-impact analysis, likely-test discovery, and budgeted context selection with decision metadata. → architecture |
| Provenance-aware memory | Durable facts can cite workspace files by content hash; changed or missing sources make those memories stale and exclude them from automatic context. |
| Intelligence routing | Task-aware model and external-agent recommendations, adjusted by prior run outcomes when evidence exists. |
| Executable DAG agents | Dependency-aware plans run ready nodes concurrently, pass predecessor results forward, block downstream failures, and isolate edit workers in Git worktrees by default. |
| Validated skills | Skill improvements can be promoted only after recorded A/B trials show uplift without regressions. |
| Cost-aware | Anthropic prompt-cache breakpoints on the stable prefix, decay- and access-aware memory ranking, and a usage_report tool that prices spend from a user-editable table — never a guessed number. |
| Scheduled work | Agent runs, direct tool calls or host shell commands on an interval, a cron expression or a one-shot timestamp. → automations |
| Extensible | Plugins register tools, skill directories, MCP servers and event listeners in-process; bridges delegate scoped work to Claude Code, Codex, OpenCode, Copilot CLI, Hermes or Aider. → extension boundary |
| Permissive by default | permissionMode: "overdrive" — host filesystem scope, no per-command approval dialogs, and recovery through automatic checkpoints and an append-only audit log. → read this first |
Requirements
- Node.js 22 or newer
- A terminal at least 80×24 (UTF-8 and truecolour are used when available, and degraded cleanly when not)
- Git and ripgrep recommended
- Any instruction-following model through Ollama or another configured provider
Install from npm
npm install -g maskshift
cd /path/to/repository && maskshiftOr run it without installing anything:
npx maskshift --workspace /path/to/repositoryOr run from the source directory
git clone https://github.com/nafeeur/MaskShift.git
cd MaskShift
./start.sh --workspace /path/to/repositoryOr install to your user account from source
./install.sh
cd /path/to/repository && maskshiftThe installer copies MaskShift to ~/.local/lib/maskshift and links ~/.local/bin/maskshift.
It does not run npm install, because there is nothing to install — MaskShift has zero runtime dependencies.
MaskShift's default model reference is ollama:auto: it discovers your installed Ollama models
and prefers the strongest coding-oriented one it finds.
# 1. Pull a coding model
ollama pull qwen3-coder:latest
# 2. Point MaskShift at Ollama and pick the model
export OLLAMA_BASE_URL=http://127.0.0.1:11434
export MASKSHIFT_MODEL=ollama:qwen3-coder:latest
# 3. Open the interface on a repository
./start.sh --workspace ~/code/my-projectType a task into the composer and press ↵. To run one task headlessly instead:
./start.sh run "Map this repository, repair the highest-impact defect, add tests, and verify." \
--workspace ~/code/my-project \
--model ollama:qwen3-coder:latestA remote Ollama host works the same way — set OLLAMA_BASE_URL=http://model-host:11434.
If a model has no native tool API, MaskShift detects it on the first request and switches to an in-prompt text protocol, so the whole harness still works. → models without native tool calling
Model references are provider:model:
ollama:auto openai:<model-id> anthropic:<model-id>
lmstudio:auto openrouter:provider/model gemini:<model-id>
vllm:auto
Cloud providers read their keys from the environment:
export OPENAI_API_KEY=... export ANTHROPIC_API_KEY=...
export OPENROUTER_API_KEY=... export GEMINI_API_KEY=...Everything else lives in ~/.maskshift/config.json — providers, MCP servers, agent bridges,
hooks, indexing, memory ranking, automations and interface preferences. Press f2 in the
interface to edit the core settings without touching the file, or use maskshift config set.
→ full configuration reference · environment variables · data locations
maskshift opens a full-screen terminal application built on a bespoke, zero-dependency
renderer, driven equally by keyboard and mouse. Eight views switch with 1–8, ctrl+b toggles
the right rail, and ctrl+k opens a fuzzy command palette over every action.
| View | Holds |
|---|---|
| 01 HEIST | Transcript and composer — markdown, syntax-tinted code, coloured diffs, live tool calls |
| 02 FILES | Workspace tree with a syntax-highlighted preview — images decode and render inline |
| 03 ARSENAL | Every native tool and skill, searchable, with parameter schemas — and x to run one yourself |
| 04 NETWORK | MCP servers: bundled, workspace-configured, or pulled live from the official registry |
| 05 MOD SHOP | Automations, plugins, agent bridges, browser profiles and background processes |
| 06 TERMINAL | The host shell |
| 07 BROWSER | A live, clickable, typeable view of a running browser tab — the same CDP connection the browser tools drive |
| 08 GIT | Working tree changes (stage/unstage/discard/commit), commit log, branches, stash, MaskShift checkpoints and worktrees — plus push/pull/fetch from anywhere in the view |
Inline images — a Files preview, a chat screenshot result, or the live BROWSER view — render
through whatever the terminal actually supports, picked automatically: the Kitty graphics
protocol, iTerm2's inline-image escape, or a universal Unicode half-block fallback for everything
else, degraded further to a clear message on a terminal with no Unicode support at all. None of
it needs a dependency: the PNG, JPEG and BMP decoders are hand-written against Node's built-in
zlib.
 03 ARSENAL — see exactly what a run can reach before it uses it |
 04 NETWORK — connect a server on demand |
 ctrl+k — every action, nothing buried behind a memorised key |
 The rail — plan, loadout telemetry, event bus |
 08 GIT — changes, log, branches, stash, checkpoints and worktrees, with push/pull/fetch from anywhere |
 05 MOD SHOP — automations, plugins, agent bridges and processes |
Below 108 columns the rail hides itself and the header sheds telemetry, so the same eight views
work in a narrow split pane. MASKSHIFT_MOUSE=off (or f2) hands text selection back to the
terminal, and NO_COLOR, MASKSHIFT_COLOR=off and MASKSHIFT_ASCII=1 each produce a clean,
aligned fallback.
→ keys, views and the design system
Everything the interface can do is also a subcommand, and every subcommand takes --json:
maskshift run "make the failing tests pass" # one headless run, streamed
maskshift tools run shell_exec '{"command":"npm test"}'
maskshift mcp registry playwright && maskshift mcp install io.github.microsoft/playwright-mcp
maskshift automation create nightly --schedule "every 6h" --prompt "Review the diff and fix regressions"
maskshift workspace search "checkpoint restore" --json | jq -r '.[].path'
maskshift doctor
maskshift mcp serve --workspace ~/code/my-project # MaskShift itself, as an MCP servermaskshift [tui] [PROMPT] open the interface (the default command)
maskshift run "PROMPT" one headless run, streamed to stdout
maskshift daemon resident automation scheduler, no interface
maskshift doctor environment and provider diagnostics
maskshift <workspace|session|tools|skills|mcp|plugins|automation|browser|config> ...
Global flags: --workspace PATH, --model REF, --config PATH, --json, --no-color.
| Document | Covers |
|---|---|
| Configuration | Every setting, provider, MCP definition, hook, automation and data location |
| Tools | The complete native tool inventory |
| Skills | Bundled skills, and where MaskShift looks for more |
| Interface | Keys, views, the design system and its rules |
| Command line | Every subcommand and flag |
| Architecture | The agent loop, lazy capability fabric, persistence and extension boundary |
| Permissive execution | What overdrive means, what stays observable, and where the limits are |
| Tool verification | Per-tool coverage and live-integration limits |
| Release verification | What the automated suite covers before a release |
CONTRIBUTING.md has the expectations changes are held to — lazy capability
loading, precise tool schemas, regression coverage, a dependency-free runtime and interface.
SECURITY.md describes the trust boundary overdrive mode assumes and where to
report a real vulnerability. examples/plugins/telemetry-pack
is a complete worked plugin to start from; the bundled plugin-authoring and skill-creator
skills cover the rest of the extension surface.
npm run check # syntax validation across every source module
npm test # unit and integration suite, including renderer tests
npm run test:tools # all-native-tool scenarios and edge-case regressions
npm run smoke # end-to-end agent run plus a full interface paint
npm run verify # check + tests + smoke
npm run docs # regenerate the tool and skill inventories
npm run capture # regenerate docs/screenshots from the real rendererThe screenshots above are produced by npm run capture through the same code path the terminal
uses, so they cannot drift from the product.
deploy/maskshift.service— user-level systemd unit runningmaskshift daemonfor scheduled automations, with no interface.Dockerfile/compose.yaml—/workspaceand/datavolumes. MaskShift is a terminal application, so attach a TTY:docker run -it, ordocker compose run --rm maskshift.
A container limits MaskShift to the files, sockets, devices and credentials mounted into it. For full host authority, run the user service directly instead.