AiSH

AiSH is an AI-native interactive shell. You describe a task in natural language; AiSH generates the shell command internally, checks its risk, executes it, and returns the result.

Generated commands and routine safety metadata are hidden during normal use.

Interaction

$ aish
AiSH
Describe what you want to do. Type :help for settings or exit to quit.
aish> find all Python files in this folder
./src/main.py
./tests/test_main.py

AiSH has two output modes:

• raw prints the command's stdout and stderr unchanged.
• natural captures stdout and stderr, asks the local model to summarize them, and preserves exact paths, file names, identifiers, numbers, URLs, and errors.

Switch modes inside AiSH:

:output raw
:output natural

Or choose the mode when starting it:

aish --raw-output
aish --natural-output
aish -n "show the five largest files here"

Natural-language output can be the default in ~/.aish/config.toml:

[output]
natural_language_output = true

Request Flow

natural-language request
        |
        v
local model generates an internal command
        |
        v
safety classifier
        |
        +-- blocked/high risk --> refuse or request confirmation
        |
        v
selected system shell executes it
        |
        +-- raw mode ---------> stdout/stderr
        |
        +-- natural mode -----> local model summary

Safe commands execute without showing the generated command. Risky commands require confirmation and describe the risk without exposing the command. --no-exec remains available only as a development/debug preview.

Run Locally

The default development backend is mock, so the UI can be tested without a model:

cargo run
cargo run -- --natural-output "list files"

Run tests:

cargo test

Use the real local model:

scripts/download-model.sh
AISH_RUNTIME=llama.cpp AISH_LLAMA_BIN=llama-cli cargo run

The default model location is ~/.aish/models/aish.gguf.

Bundle and Install

AiSH itself compiles to one native executable:

chmod +x scripts/install-local.sh
scripts/install-local.sh

This builds a release binary and installs it at ~/.local/bin/aish. Add that directory to PATH, then start the shell with:

aish

The executable is the orchestration layer. Real AI inference also needs:

• llama-cli available on PATH, or selected with AISH_LLAMA_BIN
• the GGUF model at ~/.aish/models/aish.gguf

For a single distributable environment containing AiSH and llama-cli, use Docker:

docker build -t aish .
docker run --rm -it -v aish-data:/home/aish/.aish aish

The Docker image contains the aish binary and llama-cli; the model is downloaded into the persistent aish-data volume on first use.

Source Structure

src/main.rs       CLI entry point and help
src/app.rs        interactive loop, output modes, safety and orchestration
src/runtime.rs    command generation and natural-language summarization
src/prompt.rs     command and summarization prompts/output cleaning
src/shell.rs      Bash/Zsh/PowerShell/CMD execution and output capture
src/safety.rs     command risk classification
src/config.rs     ~/.aish configuration
src/history.rs    local request/command history
src/model.rs      GGUF model setup
scripts/          local installer and model downloader
Dockerfile        bundled Linux runtime image

Safety

The model never decides execution policy. AiSH classifies every internal command as safe, risky, high-risk, or blocked. Risky operations require confirmation; extreme destructive operations remain blocked.