Early Stage Preview
Harness Assisted SHell

Treat AI like a Unix pipe, not a magic wand.

Drop ?? anywhere. Get unstuck. Keep moving.
No vendor lock-in. No magic modes.

Expect rough edges, bugs, and breaking changes.
If that sounds fun, welcome aboard.

The Magic Syntax

You don’t “enter AI mode.” You just type.

hash — 80x24
~/projects/hash $
?? find all Go files modified today
Command Output
I'll search for Go files modified today. Found 31 files: • internal/config/config.go • internal/shell/shell.go • internal/prediction/predictor.go • internal/history/store.go Mainly focused on shell, prediction, and history components.
Features

Built for Trust & Speed.

Hash isn't just a shell with an LLM. It's an ecosystem built around control, privacy, and local learning.

Agent-Agnostic

Works with anything that speaks ACP or HTTP. Claude, GPT, Ollama, Local Llama. Your choice, not ours.

Learns Locally

Hash notices how you fix errors. If you run chmod +x after permission denied 5 times, it learns to suggest it automatically.

Privacy First

Nothing leaves your machine unless you explicitly send it. Review every payload. Control every byte.

Trust Feature

Context you can actually control.

LLMs don't need all the things. They need the right things.

Press Ctrl+P to open the picker. Select exactly what goes into the prompt. Nothing else leaves your machine.

  • Auto-detects git branch & cwd
  • Visual size gauge (stay under 8KB)
  • Filter env vars manually
Context Picker3.2KB / 8KB
[x]cwd: ~/projects/api
[x]git: feature/auth
[ ]kubectl: prod-cluster
[x]history: last 5 commands
[ ]AWS_PROFILE
SPACE toggleENTER confirm

Get started in 60 seconds

1. Install Hash

Clone and build from source. (Binary releases coming soon).

2. Configure Agent

Add 3 lines to config.toml for Claude, Ollama, or Gemini.

3. Type ??

Just type ?? help and see what happens.

Read the Guide