Alpha · 1.0.0-alpha.1
Datapoint Terminal is a desktop terminal where an AI agent works in your shell, locally or over SSH. You see every command it types, and you take over just by typing.
In plain sight
It talks with you, asks when it needs a decision, and types into the same terminal you're looking at. Nothing it does happens out of view.
Before changing anything, the agent can look around with read-only commands and propose a plan: what it found, each step with its host and commands, the risky steps flagged, and how to roll back. Its steps run only once you approve it.
When there's a decision to make, the agent asks in the conversation, with choices to pick from and the one it recommends marked. It never asks for a password: it asks you to type it into the terminal, so it never reaches the model.
Every host carries an environment label, and production stands out in red. In Ask, the default mode, each command waits for your approval, with the terminal's name and label in plain view. You can edit a command before it runs.
Save a task as a skill, then roll it out to the hosts that match its labels: one at a time or in batches, with a health check after each, stopping at a failure if the skill says so. The application enforces the rollout, not the model.
The agent keeps a document and a maintenance log for each host, as plain Markdown files in a workspace folder your team already shares, on a network drive, OneDrive or Git.
Use Anthropic or OpenAI with your own API key, or an open-weight model you run yourself, with Ollama, LM Studio or vLLM. Mark a model as local, and the application shows that the conversation stays on your machine or network.
A good terminal first
zsh, bash, fish and PowerShell, locally or on your hosts, with jump hosts, port forwarding and file transfer.
Split right or down, and search the scrollback.
Follows your system, or the theme you choose.
Know when the agent finishes, or waits for your approval or an answer.
Played back in the application, kept on your computer, also playable with asciinema.
See what each conversation uses with your provider and about what it costs, and set a spending limit.
Safety and privacy
The agent runs real commands on real machines. These safeguards make its work visible and reviewable, but none of them can make it harmless.
The terminal is never locked. Type anything and the agent stops; your keystroke still reaches the shell, and anything running is left alone.
Terminal output, the files the agent reads and your messages go to the AI provider you choose, a third party unless you run the model yourself. You can see exactly what was sent at each step.
Secrets the application recognises are masked before anything is sent, and files it shouldn't read can be listed in .terminalignore. Detection can miss a password written in prose.
Terminal's README sets out habits that keep you in control, such as starting in Ask mode, labelling production hosts and staying at the keyboard, and the known risks: what the safeguards can't catch.
Download
Version 1.0.0-alpha.1: an early test build. Expect rough edges, and anything may change before version 1.0.
The downloads appear here as soon as version 1.0.0-alpha.1 is published.
Licence
Free
Use, change and share it under the GNU Affero General Public License, version 3.
Let's talk
For organisations that can't accept the AGPL.
Yes. Datapoint Terminal is free software under the AGPL-3.0. You pay only your AI provider, directly, for what the agent uses, or nothing if you run an open-weight model yourself. Organisations that can't accept the AGPL can buy a commercial licence.
What the agent works with goes to the AI provider you choose: your messages, terminal output while it works on a task, and the files it reads, with the secrets the application recognises masked. With an open-weight model you run yourself, it stays on your machine or network. Datapoint receives none of it.
Anthropic's and OpenAI's models, with your own API key, and open-weight models served over an OpenAI-compatible API, such as Ollama, LM Studio or vLLM. A model must support tool calling; choosing one that hasn't been tested shows a warning.
No tool can promise that. Datapoint Terminal is built so you see everything the agent does and can stop it at any moment: Ask mode, the default, waits for your approval before each command, and plans show their steps and how to roll back. Using it safely lists the habits that help, and the known risks what these safeguards can't catch.
It's built for macOS, Linux and Windows, but so far it's only been used on macOS. The Linux and Windows builds of this alpha haven't been tried, and some features may not work there yet.
Not yet. Updates will be signed and checked before they're installed, and until that's in place, the application doesn't check for them. Download new versions from this page.
Free and open source, for macOS, Linux and Windows.