maxsignalhq/Sakshi

The witness for your AI coding agents.
Sakshi (साक्षी, Sanskrit for "the witness") runs and supervises AI coding agents from one terminal. It supports Claude Code and Codex. See what every agent is doing, get notified when one needs you, and know at a glance whether its work can be trusted.
An agent is asked to fix failing tests. It skips the test, deletes the assertions and lowers the coverage gate, then reports "All tests pass". Sakshi turns the session red and says exactly why. (Recorded with a scripted stand-in agent so it is reproducible; see scripts/demo .)
> Status: 0.4, early. It works and is tested on macOS, Linux and Windows, but it is young, and the trust rules are heuristics (see TRUST.md).
## Why
Running agents is easy. Knowing what they did is not. Sakshi watches each session and tells you when something looks off: tests skipped or deleted, assertions weakened, sensitive files touched, destructive commands run.
## Install
Homebrew (macOS and Linux):
sh
brew install maxsignalhq/tap/sakshi
Scoop (Windows):
powershell
scoop bucket add maxsignalhq https://github.com/maxsignalhq/scoop-bucket
scoop install sakshi
macOS and Linux, install script:
sh
curl -fsSL https://raw.githubusercontent.com/maxsignalhq/Sakshi/main/install.sh | sh
Windows (PowerShell):
powershell
irm https://raw.githubusercontent.com/maxsignalhq/Sakshi/main/install.ps1 | iex
Both scripts download the build for your machine from the latest GitHub release, check its SHA-256 against the published checksum, and install sakshi to ~/.local/bin (Windows: %LOCALAPPDATA%\Programs\sakshi, added to your user PATH). Read them first if you like; they are short. Set SAKSHI_VERSION=v0.1.0 to pin a version or SAKSHI_INSTALL_DIR to choose the folder.
By hand: release archives for macOS (Apple Silicon and Intel), Linux (x86-64 and ARM64, glibc) and Windows (x86-64 and ARM64) are attached to each GitHub release with a SHA256SUMS file. Unpack one and put sakshi on your PATH.
> macOS note: release binaries are not yet notarized by Apple. The install script is unaffected, but if you download an archive with a browser, macOS may refuse to open sakshi. Clear the download flag once with xattr -d com.apple.quarantine ./sakshi.
From source, with a recent stable Rust:
sh
cargo install --git https://github.com/maxsignalhq/Sakshi sakshi
The Homebrew, Scoop and winget manifests are generated from templates in packaging/. The winget package is not published yet. Publishing a release to the package managers is a manual step, described in docs/RELEASING.md.
Optional: git on your PATH enables change tracking, and desktop notifications use your OS's own notification service.
## Use
sh
sakshi # start your shell inside Sakshi
sakshi claude # start Claude Code inside Sakshi
sakshi claude --resume # arguments after the program go to the program
Sakshi shows a session list on the left and the selected session's terminal beside it. Agents run exactly as they would in any terminal; Sakshi only watches.
Attached (typing into a session): everything goes to the program. The one key Sakshi keeps is Ctrl+], which detaches you back to the list. Shift+PgUp / Shift+PgDn scroll back through the session's output; typing or pasting returns you to the live screen. Pasting, including several lines at once, works: programs that ask to be told about pastes get them marked as pastes, so a pasted newline is text and not Enter.