The sill tool gives AI coding agents a Bash shell on Windows

A developer released sill, a Bash shell for AI coding agents on Windows. A shell is the software that operates typed commands. Agents are made for Linux behavior, and sill gives it on Windows with no Git, no MSYS2 and no Linux layer. In a test by the maker on a GitHub test system, Codex made 0 failed calls out of 52 with sill.

How it works

sill is a Bash interpreter made in Rust. It is a changed copy of another Bash shell. It corrects the points where agent commands stop on Windows: quotes, paths, text encoding, console windows that flash up, and background servers that hang. It also operates as an MCP server, a standard method to give tools to a model.

The numbers

The maker did 59 test cases and all went through. In 18 agent tasks with a low cost DeepSeek model, the counts of done tasks were close. The difference was in failed calls and time. Bash itself gives the same result as Git Bash in 2,434 of 2,932 tests (83%).

What is not known

The maker did all the tests, and no other group has done them again. The tests used Windows 11. The GitHub test used each task one time. Most of the other Bash tests are about Unix signals and job control, which Windows does not have.

Insight: The layer between agent and machine is now a market

Akamai seals agent code in a sandbox. sill puts each agent command in a Windows Job Object, and a timeout ends all processes of the command. Akamai gave no test results. The 2 tools are between the agent and the machine, and maybe the maker of this layer sets the limits of agents. A shell that makes agents work does not limit them. Builders also want a sandbox.