Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

pmndre/pine

AI Engineering

A terminal-based agent harness where agents subscribe to events and act on schedules, with coding and computer-use tools.

232
4
TypeScript
MIT
5 days old

What it is

Pine is a TypeScript agent harness that runs in the terminal, built around the idea that agents should react to events rather than only answer chat messages. Event sources can include folders, commands, webhooks, Telegram, MQTT, FTP, web addresses and schedules, and agents wake up when one of them fires README. It also includes what the README calls the basic stuff: chat, coding tools, memory, git diffs, undo, MCP servers and skills.

Agents can delegate to or create sub-agents in parallel, and a pinned tab keeps an agent and its task around across sessions. Configuration is largely done by asking the main agent to create agents, MCPs, event sources, skills and schedules through tool calls, instead of editing JSON by hand setup. A built-in BM25 manual lets agents look up API details on demand.

Quick start

You need macOS or Linux, Node.js and npm. Install the CLI globally, then initialize it:

npm install -g @pmndre/pine
pine init

pine init sets up the browser used by the web tools and the ~/.pine user folder; on Linux it also installs the sandbox used in safe mode and asks before doing so. Check the setup with pine check, then open the terminal UI:

pine check
pine

If no daemon is running, pine starts one inside the UI process. To keep agents reacting without an open terminal, run the daemon as a service with pine start operations.

Connect a model provider next, for example a subscription or a local model server:

pine login opencode-go
pine model add studio http://127.0.0.1:8080

Switch models with /model in the UI. To work on a project, start Pine inside that folder so the agent gets it for the conversation:

pine .

The code pane (ctrl+g) shows git changes, files and tool output, and /undo reverts your last message along with the files it changed coding.

Docker and Windows

Windows is not a first-class target. The README suggests running Pine in a container instead Docker:

docker run -it -v pine:/root/.pine node:24 bash

Inside the container, install the package and run pine init as usual. Agents default to auto mode in this setup. If you want safe mode with a per-agent sandbox inside the container, the container needs additional capabilities (SYS_ADMIN, NET_ADMIN) and relaxed seccomp/AppArmor profiles, which materially weaken Docker's own isolation.

Constraints

  • Platform support is macOS and Linux; complete Windows support is explicitly unresolved.
  • Computer use is macOS-only, so that tool set is not available on Linux.
  • Safe mode relies on a sandbox (bubblewrap on Linux); in containers the sandbox requires broad capabilities that reduce container isolation.
  • The harness can run commands, edit files, browse the web and control apps, so the security model and what is not protected matter before giving it broad access security.
  • Model access is by subscription logins or model server addresses; Pine does not ship its own model.
  • Licensed under MIT.

Documentation map

The docs start at docs/SKILL.md, which indexes overview, TUI, coding, operations, configuration, setup, agents, events, event sources and source types. The terminal UI is documented in tui, event routing and schedules in events, the available source types in source-types, and MCP integration in mcp.

HotRepo's take

The project appeared on HotRepo's new-project board for TypeScript on 2026-11-16, gaining 13 stars in a roughly 4.5-hour measured window and reaching 232 stars overall with 4 forks. That is a small but real early signal: the star count is high relative to forks, and there is no GitHub all-language daily rank recorded yet. The design direction — event-driven agents that can drive folders, webhooks and even physical devices through MCP — is the part worth watching; whether the agent-managed configuration and terminal UI hold up under real use cannot be inferred from the listing data.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. pmndrs/pineA terminal-based agent harness where agents subscribe to events, schedules, and MCP tools and react autonomously.TypeScript117
  2. Balanov/shimA tiny local-first harness for running terminal AI agents with any model, MCP servers, and Agent Skills.Go117
  3. deepseek-ai/deepseek-harnessAn open-source agent harness from DeepSeek AI built on an everything-is-a-plugin architecture for running configurable AI agents.TypeScript274,658
  4. KKKKhazix/agent-autoloopSelf-hosted automation hub that runs command-line coding agents on schedules and events, pausing for your approval when needed.TypeScript3,225