← Back to archive

Repo of the Day

pmndre/pine: a reactive harness for agents

Published: Nov 18, 2026

Open repository ↗

a reactive harness for agents. Contribute to pmndre/pine development by creating an account on GitHub.

Summary

PINE is a TypeScript-based harness for building AI agents that act in response to events rather than waiting only for chat input. It combines an event bus, scheduled tasks, coding tools, model connections, delegation, and terminal or desktop interfaces in one system.

What it is useful for

PINE is useful when an engineering workflow should continue without someone keeping a chat window open. An agent can subscribe to folder changes, webhooks, MQTT or FTP events, website activity, commands, and other sources. For example, it could watch a scans directory and extract summaries and payment details from new documents, or react to a Slack message by interpreting its contents.

It also supports more conventional agent work. Teams can use it for chat, task execution, and orchestration, while agents can delegate work to other agents or create parallel sub-agents with distinct prompts. Built-in coding support includes git diffs, file inspection, tool output, and /undo for taking back the last message and its file changes.

The project’s main design choice is event-driven configuration: agents can be told what to watch, and they can create agents, skills, MCP connections, event sources, and schedules through administrative tool calls instead of requiring users to edit configuration files manually. The README also describes a terminal UI with tabs, inline media, and drag-and-drop, plus a macOS desktop app.

How engineers can use it

Install the CLI as documented:

npm install -g @pmndre/pine
pine init

pine init prepares the environment, including the browser used by web tools and the ~/.pine user directory. On Linux, it also sets up the sandbox used by agents in safe mode. Validate the installation with:

pine check

For project work, start PINE from the repository directory:

pine .

The agent receives that folder as its working context. In the terminal UI, ctrl+g opens the code pane with git changes, diffs, files, and tool-call output. This makes PINE applicable to tasks such as implementing a change, running tests, and reviewing the resulting diff.

To keep agents reacting when no terminal is open, run it as a service:

pine start

Models can be connected with pine login opencode-go, pine login anthropic, or pine login openai-codex; a local model server can be added with pine model add studio http://127.0.0.1:8080.

PINE natively supports macOS and Linux. Windows users can try the documented Docker path, although the README says full native Windows requirements remain unsettled. The README points to docs/SKILL.md as the documentation index and docs/references/security.md for sandbox and secret-handling details.