Manual tool juggling
You memorize flags for a dozen CLIs and translate intent into syntax in your head, every single time.
Replace bash with intent
aish is an AI-native shell — no POSIX layer underneath, just a stateful agent loop. Describe what you want; it understands, calls the right tools over MCP, reads and writes files, and runs your programs.
The cost of context-switching
Your work lives in five tools that don't talk to each other.
A terminal for commands. An IDE for code. Dashboards for infra. Chat for the team. Every switch is manual, and the thread of what you were actually trying to do gets dropped on the floor.
Lost context
Each window is a fresh start. The reasoning behind a chain of commands evaporates the moment you alt-tab.
Flow, interrupted
The constant switching tax is paid in minutes — and in the deep focus you never quite get back.
One conversational surface
Intent in. Results out. Nothing in between.
aish is a single stateful agent that understands what you mean, picks the right tools, and streams the work back to you — so you stay in flow.
Natural-language intent
Stop translating goals into syntax. Say “tail the error logs for the api service over the last hour” and aish does the parsing.
MCP-native tools
File ops, git, HTTP, cloud APIs — anything exposed over the Model Context Protocol becomes a first-class capability in your shell.
Stateful agent loop
aish remembers the session — env, jobs, what just happened — and reasons across steps instead of forgetting between commands.
Background coordinators
Hand off long, parallel work to background agents. They run, report, and deliver results back into your session when done.
Local & offline models new
Run fully offline with in-process GGUF inference — no API key, no network. Qwen3-1.7B loads on first use; point AISH_LOCAL_MODEL_ID at any GGUF model with no rebuild.
Switch backends live
Hybrid brain: lean on Claude for deep reasoning, then :backend local to go private and offline mid-session. One shell, your choice of cloud or local — per task.
Where it stands
A different kind of shell.
aish trades the POSIX layer for an agent loop. Here's how that lands against the shells you know.
| Capability | aish | bash | zsh | fish |
|---|---|---|---|---|
| Natural-language intent | ||||
| Stateful agents | ||||
| MCP tool integration | ||||
| No POSIX shell required | ||||
| Works fully offline |
Offline local-model inference shipped in v0.19 — switch with :backend local and run any GGUF model in-process, no API key required.
Where it's going
Shipping toward 1.0.
Offline-capable core today; durable agent state and team-scaled shells next.
-
v0.19 shipping now
Local & offline models — in-process GGUF inference (mistral.rs, Qwen3 by default), tool calling from local models, and live
:backendswitching. Plus the core: REPL, MCP tools, background coordinators, sessions, and skills. -
v0.20 next
Agent & skill registry, durable agent state across restarts, and tougher error recovery in long-running agent loops.
-
v0.21 later
Team-scoped shells with shared, auditable workflows and synced skills across a team.
-
v1.0 milestone
Production stability guarantee, Kubernetes / cloud-native runners, enterprise features — SAML and audit logs.
Replace bash with intent.
Clone, build, and open the REPL. You'll be talking to your shell in under two minutes.
git clone https://github.com/LightHeart-Ventures/aish
cd aish && cargo build --release
./target/release/aish