Vis

Vis is a coding agent you can adapt to your tools and workflow. Give it tools for your project, measure the results on real tasks, and use failures to improve those tools and checks.

Repeat the same tasks after each change to compare the results.

Follow the work from your terminal, browser, desktop or phone, with the same sessions on each.

Install

Install Vis on the computer where your projects live. The installer does not need Java or Git. Native packages support Apple silicon macOS and Linux (x64 or ARM64). On Windows, use Linux in WSL2. See Runtime distributions for source builds and other installation options.

The installer writes to ~/.local/bin and can update your shell profile. You can read it first.

curl -fsSL https://github.com/Blockether/vis/releases/download/installer/install-vis-agent | bash

This installs the engine, Python and two of the three clients. All clients show the same sessions, so you can start a task in one client and continue it in another.

Client	What it is	How to get it
Terminal client (TUI)	Vis in your terminal, in the project directory.	Included. Run vis-agent tui.
Web app	Vis in a browser, on this computer or on other devices on your network.	Included. Run vis-agent web.
Companion app	Vis on your desktop, phone or tablet, connected to the engine on your computer.	Run vis-agent desktop on macOS or Linux to download and open it automatically. Or install the desktop app: Linux ARM64 · All packages · Desktop setup Phone and tablet (public beta): Connect the desktop app · Pair a phone

Rationale

A capable model can write code, but it does not know how your team works. It does not know which tests matter, how changes get reviewed or what to check before a release. You do. Vis lets you put that knowledge into tools the agent can use, instead of a growing list of reminders.

For example, a function can select and run the right tests for a change. A hook can check the code after an edit and report problems. You define the operations and checks. The model decides how to combine them. Your instructions explain the process, while your code carries out its repeatable parts.

Python connects those steps. Models already use it to work with code and data. Vis lets them use the same language to compose your tools, inspect results and reuse useful helper functions. You can start with the built-in tools and add extensions as you need them.

You can also see how the answer was reached. Activities show actions, results and failures in the conversation. Your extensions choose what to show, so you can read a test result or build summary without deciphering a stream of shell commands. Switch devices to follow the same work, rather than start again.

Read the full Rationale in the manual.

Documentation

Read the manual at vis.blockether.com or in resources/vis-docs/. Its five modules, in reading order:

• Intro: Rationale and Getting started.
• Concepts: one page for each feature, such as Sessions.
• Programmatic access: Python SDK, Python API reference and HTTP API.
• Extensions: Extending Vis, code checks example and the Extension Center.
• Reference: exact facts, such as Keyboard shortcuts.

Screenshot gallery

All images come from one demo project, fieldnotes, after real tasks: a project tour, a bug fix with a regression test and a release preparation.

License

Apache License 2.0 — see LICENSE.

The speech service can install third-party models with separate license terms. THIRD_PARTY_MODELS.md lists their authors and licenses.