Open-source tools for agent-powered work

An agent that does the work.A terminal that keeps watch.

miaoAI agentmacOS · Linux · Windows

Writes code, researches topics, manages files and automates tasks — with the models you choose and a running tally of what each turn costs.

macOS / Linux

curl -fsSL https://mtty.dev/miao/install | bash

Windows (PowerShell)

irm https://mtty.dev/miao/install.ps1 | iex

mttyTerminal · editor · agentsmacOS · Linux · Windows · preview

An AI-native terminal and editor in Rust: terminals, files, remote hosts and agents in one window — and it knows when an agent is working, waiting on you, or done.

Download the preview

Current miao example session with an inline edit diff, response and context telemetry. — **miao**AI agent

miao model provider selection — **miao**AI agent

01 the products

Use either one. They are separate projects.

miaoAI agent

Your models. Your workflow. Less wasted context.

Steer a running task: Add a requirement mid-task. It is saved first, then folded in at the next safe turn boundary.
Delegate with its own context: Hand scoped work to a subagent; sessions can find and message each other.
See what every turn costs: Tokens, estimated cost, time to first token and prompt-cache hits, per turn.
Keep the record: Sessions are durable: reopen, fork and export them instead of losing a window.

TypeScript · Rust accelerators · MITExplore miao →

mttyTerminal · editor · agents

Terminal, editor and agents in one native window.

Rendered on the GPU: Smooth terminal rendering with damage-driven redraws and measured performance budgets.
A workspace, not just a window: Tabs, recursive splits, quick file navigation and native editor panes with syntax highlighting, language servers and live previews.
Aware of agents: Pane badges, notifications, a sleep guard and a prompt queue driven by agent state.
Remote hosts like local folders: A host library, keys, SFTP, port forwarding, jump hosts and snippets over your own OpenSSH.

Rust · Apache-2.0 · pre-releaseExplore mtty →

02 better together

Run four agents. Read none of their windows.

Inside an mtty pane, miao reports its state over the terminal's control plane. mtty badges the pane, raises a notification when an agent needs you, keeps the machine awake while one is working, and sends your queued prompt the moment it goes idle.

Neither needs the other. Outside mtty the report is a no-op; with no agent reporting, mtty is simply a terminal. mtty also ships state hooks for other agent CLIs.

03 in pictures

Terminals, editor and agents in one window.

An illustrative session shows the edit diff and context telemetry; model selection stays in the same conversation.

miao · review, then choose a modelAn illustrative session shows the edit diff and context telemetry; model selection stays in the same conversation.View full size ↗

mtty · agent state, liveControl-plane events update the tab badge and Agent panel through processing, awaiting, error and idle.View full size ↗

miao · bring your own providerBrowse the provider picker and choose the service that fits your work.View full size ↗

mtty · editor and live previewA Markdown edit updates the adjacent preview, with Mermaid, tables and code blocks.View full size ↗

miao · the browser workspaceExpand a code diff and open the review panel in the browser, using the same illustrative session.View full size ↗

mtty · inline edit reviewReview a proposed change in the editor, then accept it through the command palette.View full size ↗

**mtty · local and remote**A grouped host library with file access and saved port forwards; all addresses shown are examples.View full size ↗

**mtty · recursive splits**Git changes and a running local HTTP server side by side in the same workspace.View full size ↗

04 status

Early, moving fast, and plain about it.

01
Pre-1.0
Both projects release often and change between versions. Release notes live on GitHub.
02
Docs from the source
User guides for both products are synced from their repositories; each synced page names its source commit.
03
Claims with caveats
Benchmarks say what they measure. Features that are opt-in or experimental say so.

Start with one command.