kaish (会sh)

A predictable shell for AI agents. A distilled Bourne-like shell with builtin core utilities. Kaish is embeddable, has stronger validation ahead of execution, and can embed MCPs as commands.

Install

cargo install kaish

This is preferred for now while kaish is still experimental. Containers and binaries are in future plans when things stabilize a bit more and I have time (or PRs!).

Why kaish?

Traditional shells have evolved syntax with many sharp edges. kaish implements the commonly-used parts of sh while eliminating entire classes of bugs at the language level:

• No implicit word splitting — $VAR is always one value, never split on spaces
• No glob expansion — tools handle their own patterns, or use glob builtin
• Structured iteration — for i in $(seq 1 5) works via structured data, not word splitting
• Explicit splitting — use split "$VAR" when you actually need word splitting
• No backticks — only $(cmd) substitution
• Strict booleans — TRUE and yes are errors, not truthy
• Pre-validation — catch errors before execution, not at runtime

Skills transfer from bash. Footguns (hopefully) don't.

Quick Tour

#!/usr/bin/env kaish

# Familiar bash-style syntax
GREETING="Hello"
echo "$GREETING, world!"

# Control flow
if [[ -f config.json ]]; then
    echo "Config found"
fi

# For loops - no implicit word splitting!
for item in one two three; do      # literal items
    echo "Processing: $item"
done

for i in $(seq 1 3); do            # structured data iteration
    echo "Count: $i"
done

for file in $(glob "*.txt"); do    # structured data iteration
    echo "Found: $file"
done

# Pipes and redirects
cat urls.txt | grep "https" | head -n 10 > filtered.txt

# Expand glob patterns
glob "**/*.rs" --exclude="*_test.rs"

# Parallel execution with scatter/gather
seq 1 10 | scatter as=N limit=4 | echo "processing $N" | gather

Language Features

See Language Reference for complete syntax. Use help builtins or help <tool> for per-tool docs.

Builtins

kaish builtins run in-process — no subprocesses, no PATH lookups, no platform variance. They exist because agents need tools they can verify: a grep that behaves identically everywhere, a jq that always uses the same filter syntax, an awk that never surprises.

Design principles:

• Verifiable — each builtin has a schema (params, types, examples) exposed via help <tool>. Agents can introspect before calling.
• Convention-following — flags and behavior match the patterns deeply embedded in training data and decades of existing scripts. grep -rn, sed 's/old/new/g', awk '{print $1}' all work as expected.
• 80/20 — implement the features used 80% of the time, deliberately omit the 20% that add complexity without proportional value. Missing features compose via pipes.
• ERE everywhere — all regex uses Extended Regular Expressions. No BRE/ERE confusion.

Components

kaish is built as a set of crates that can be used independently:

kaish-kernel

The core execution engine. Lexer, parser, interpreter, builtins, VFS.

use kaish_kernel::{Kernel, KernelConfig};

let kernel = Kernel::new(KernelConfig::default())?;
let result = kernel.execute("echo hello | tr a-z A-Z").await?;
println!("{}", result.out);  // "HELLO"

The kernel is embeddable — no external dependencies, no subprocess spawning for builtins.

kaish-repl

Interactive shell with readline support, history, and tab completion.

$ kaish
kaish> for f in $(glob "*.rs"); do wc -l "$f"; done
  142 main.rs
   87 lib.rs
kaish>

kaish-mcp

MCP server exposing kaish as tools for AI agents. Builtins produce structured data internally — humans see clean readable text. The --json flag is still available as a rendering option but your agent will see json either way.

Installation

Add to your MCP client configuration:

{
  "mcpServers": {
    "kaish": {
      "command": "kaish-mcp"
    }
  }
}

Tools

execute — Run kaish scripts in a fresh, isolated environment.

Supports: pipes, redirects, here-docs, if/for/while, functions, builtins,
${VAR:-default}, $((arithmetic)), scatter/gather parallelism.

NOT supported: process substitution <(), backticks, eval, aliases.

Paths: Native paths work within $HOME (e.g., /home/user/src/project). /tmp for temp files.

Output is clean text by default — simple commands return plain text, structured builtins (ls, mounts, vars) return readable tab-separated values. Use --json on any command for structured JSON output when needed.

help — Discover syntax, builtins, VFS, and capabilities.

Topics: overview, syntax, builtins, vfs, scatter, limits
Tool help: help grep, help jq, help git

MCP Client Mode

kaish can also consume external MCP tools, appearing as namespaced commands:

# External MCP tools look like CLI commands
exa:web_search --query "rust parser combinators"

# Pipe MCP results through kaish builtins
github:list_issues --repo="foo/bar" | jq '.[] | .title'

Why an MCP shell?

AI agents need to compose operations — filter outputs, transform data, iterate over results. Raw MCP tools are individual operations; kaish lets agents combine them:

# Filter and transform in one script
ls src/ | grep "\.rs$" | head -n 5

# Iterate over results
for f in $(glob "*.json"); do
    jq ".name" "$f"
done

# Parallel processing
seq 1 10 | scatter as=N limit=4 | echo "processing $N" | gather

The kernel runs builtins in-process (no fork/exec), making it fast and predictable.

Why 会sh (kaish)?

会sh was originally prototyped as part of 会術 Kaijutsu and was separate enough it made sense to split it out. Amy was also a fan of ksh and pdksh back in the 00s so k-ai-sh seems fun.

Building from Source

git clone https://github.com/tobert/kaish
cd kaish
cargo build --release

Contributing

Agent-generated PRs are welcome! 🤖 This project is built with AI agents and we love seeing what other agents come up with. That said, please have your agent (or another model) review the PR before submitting — a few tokens on review goes a long way. Same goes for issues: agent-filed is fine, just make sure it makes sense.

If you're working with AI coding agents, you might also be interested in:

• gpal — Gemini as an MCP server (pairs well with Claude Code)
• cpal — Claude as an MCP server (pairs well with Gemini CLI)

License

MIT