Bash on Windows without the VM: fauxnix translates instead of emulating

What it does

fauxnix parses bash commands into an AST, translates them into PowerShell scripts, executes them natively, and returns output that looks like GNU/Linux: ls -l columns, bash-style error messages, coreutils exit codes. It ships as an MCP server plus a CLI, with one-command setup for Claude Code, Codex, OpenCode, Kimi Code, and Qwen Code. No VM, no WSL, no bash toolchain — just Node and the PowerShell that’s already on the box.

The interesting bit

The core insight is that LLM agents are dramatically better at bash than PowerShell — bash dominates training data, so models on Windows produce “looks right, doesn’t run” commands. Instead of emulating a Linux environment, fauxnix translates a curated subset (109 commands) and fails loudly on what it can’t handle, so the agent never gets silently-wrong results. Their own benchmarks show the gap: one model was 3.1× slower with 24 error events writing raw PowerShell versus zero errors through fauxnix.

Key highlights

• 109 translated commands, output-matched against real GNU coreutils via a 253-case differential corpus (Git Bash is literally the test oracle)
• Zero LLM calls at runtime — deterministic translation, 400+ automated tests
• MCP session persists cwd, env vars, export/unset, positional parameters across tool calls — behaves like a logged-in shell, not stateless exec
• bash_batch compiles multi-step plans up front and runs them atomically in one MCP round trip
• Unknown commands (git, node, npm…) pass through natively; UTF-8/GBK encoding traps and CRLF handled at the process boundary
• Experimental facade mode implements the bash process contract, aiming to replace an MCP sidecar with a fauxnix-backed bash.exe

Caveats

• Windows-only by design; requires PowerShell 5.1+ and Node ≥ 18 (PowerShell 7 is opt-in)
• The npm package is fauxnix-cli, not fauxnix — the latter belongs to an unrelated 2015 websocket library
• sed/awk scripts are parsed, so unsupported constructs throw named errors rather than silently misbehaving — better than wrong output, but still a wall
• The README itself says it plainly: if Git Bash already works for you, keep it

Verdict

Worth a look if you run AI agents on Windows — especially agent fleets, CI runners, or locked-down machines where shipping a bash toolchain is painful. If you’re happily on Git Bash or WSL, there’s little reason to switch.