by pmndrs
Pine review, pricing and verdict
Open-source agent daemon whose agents wake on events, schedules and chat
- Best for
- Developers who run agents on a home server or workstation
- Not for
- Windows-only office worker
Should you use Pine?
Choose Pine if you run macOS or Linux with Node 24 and want agents that react to events and schedules instead of waiting for chat. It is still a 0.x release with breaking changes between versions and a very small community, so it suits tinkerers, not teams that need support.
Agents are woken by events and schedules, not only by chat, inside an OS sandbox.
The pmndrs name and an event-first design set it apart, but a two-day-old project with no hosted tier has little lock-in and rivals can copy the idea.
Also for
- Tinkerers wiring folders, devices and chat to agents
- Engineers who want sandboxed, approval-controlled automation they can self-host
Show 2 more
- Developer running a home automation server
- Platform engineer prototyping event-driven agents
Also not for
- Team that needs a vendor support contract
Main trade-off
Very new, macOS and Linux only, with breaking changes between releases.
Pine pricing
Free and open source under the MIT licence, with no paid plan or hosted version found. You pay your model provider, whether a Claude or ChatGPT subscription, a hosted API or your own hardware for a local model server.
Before you commit
- The model provider bill or local hardware, since Pine ships no model
What Pine does on its own
- Scope of autonomy
- Semi-autonomous
- Human approval boundary
- Human in the loop
- Maturity
- Beta
Safety controls
- Runs sandboxedSupported
- Audit loggingSupported
- Human in the loopSupported
- Output filteringSupported
What it can operate
- Web browsing
- File system
- Shell commands
- Code execution
- API calls
Where Pine runs
- Deployment
- Local
- Where it runs
- CLI · API
- Memory
- Long term
Models it runs on
- Any tool-calling model through the pi SDK: Anthropic, OpenAI Codex and OpenCode Go sign-ins, or a local llama-server
Systems it connects to
| System | Type |
|---|---|
| Telegram | Native |
| MQTT | Native |
| Model Context Protocol servers | Plugin |
| Webhooks | Webhook |
What Pine does
Pine is an MIT-licensed, event-driven multi-agent daemon with a terminal UI from the pmndrs collective, first published on 10 November 2026. Agents wake on events such as new files, Telegram messages, MQTT topics and webhooks, run inside an operating-system sandbox, and use MCP servers, skills and sub-agents with whichever model you connect.
Event-driven agents
Agents subscribe to sources in the real world and on the network, and wake when something happens there, with no chat message needed.
Schedules with a task
An agent can be woken by cron, every-so-often or one-off schedules that carry a task.
Set up by asking
Agents add MCP servers, skills, sources and schedules through tool calls, so there are no configuration files to edit by hand.
Sandbox and three modes
Shell and file tools run in an OS sandbox, and each agent is in manual, auto or bypass mode, from approving every change to checking nothing.
Show all 7 features
Delegation and sub-agents
The main agent creates other agents, hands them work, and starts parallel sub-agents with their own prompts.
Terminal UI like a desktop app
Tabs, buttons, inline images and video, and drag-and-drop with thumbnails, on Kitty-protocol terminals such as Ghostty.
Self-discovery manual
A built-in search tool lets an agent look up how pine works on demand instead of carrying the docs in its context.
Strengths and limits
Strengths
- Events and schedules wake agents with no prompt, so a README example such as watching a scans folder and extracting payment details from each new file runs unattended.
- Shell and file tools run inside an operating-system sandbox (sandbox-exec on macOS, bubblewrap on Linux), and the security docs lay out each layer and what is not solved.
- It is free under MIT with no hosted tier, and works with a Claude or ChatGPT subscription or a local model server, so no extra vendor bill is added.
Limits
- It is days old: the repository was created on 10 November 2026, the npm package is still a 0.x release, and commits marked as breaking changes shipped between releases, so behaviour may change under you.
- Only macOS and Linux are supported; the README says Windows users can try Docker and that what Windows needs is still open.
- Its own security page lists what is not solved: the reviewer is itself a model that can be wrong, local MCP servers run unsandboxed, and bypass mode acts with your rights on your files.
- The community is tiny, with 50 stars and no forks on GitHub when read on 12 November 2026, and the issue tracker is the only support channel found.
Pine alternatives
| Agent | Where it runs | Price | Maturity | Rating |
|---|---|---|---|---|
| Pine This page | Your machine | Free | Beta | 3.3 |
 OpenClawPick Pine if you want agents woken by folders, MQTT, webhooks and schedules plus a terminal UI; pick OpenClaw if you want one self-hosted gateway behind your chat apps. | Cloud · Your machine · API | Free | Beta | -- |
 Hermes AgentPick Pine for event-driven agents in an OS sandbox on macOS or Linux; pick Hermes Agent for a model-agnostic, self-hosted agent with optional cloud hosting and a much larger community. | Cloud · Your machine · API | Free plan | Available | 3.7 |
 NanoClawPick Pine for event sources, schedules and sub-agents under an OS sandbox; pick NanoClaw for one Docker container per session behind your chat apps. | Your machine | Free | Available | -- |
Also compared
OpenCodePick Pine for agents that react to events and schedules; pick OpenCode for a coding agent in the terminal, desktop and IDE that works with any model provider.
Pine facts
- Works offline
- Yes
- Minimum skill level
- Intermediate
- Budget fit
- Free
- Open source
- Yes
- Ships
- Daily
- Last major update
- Maker founded
- 2026
What changed
Questions about Pine
How much does Pine cost?
Pine itself is free: it is released under the MIT licence and installs with one npm command. The cost comes from the model you connect, which can be a Claude or ChatGPT subscription, a hosted provider such as OpenCode Go, or a local llama-server that costs only your own hardware. Hokai found no paid Pine plan, hosted version or usage fee in the repository or its docs.
Is Pine open source and free to run?
Yes, the source is public on GitHub under the MIT licence and there is no paid tier. You need Node 24 or newer, macOS or Linux, and a model that supports tool calls. On Linux the agent sandbox needs bubblewrap, and without it agents get no shell.
What does Pine do without asking?
Each agent runs in one of three modes. The default, auto, asks almost nothing: only your approve list asks, and after an agent reads outside content a second model call, the reviewer, checks actions that send, reach out or change its memory. Manual mode waits for you on every change, and bypass checks nothing.
How do I connect a model to Pine?
Run pine login with a provider, for example pine login anthropic or pine login openai-codex for a Claude or ChatGPT subscription, or pine login opencode-go, then choose the model with /model in the UI. A model server of your own is added by a command or by the main agent, with its key read from the .env file, and the docs say to start a local llama-server with tool-call support.
What can wake a Pine agent?
Agents subscribe to event sources: Telegram, MQTT, watched folders, webhooks, FTP uploads, polled web addresses and repeated commands. Schedules can also wake an agent by cron, interval or at one set time. An agent sets a source up when you ask it, and only the main agent can add a new kind of source.
About Pine
Pine is an open-source, event-driven agent daemon from the pmndrs collective, first published on GitHub on 10 November 2026. Its README sums it up as "React for AI": instead of waiting for a chat message, an agent subscribes to things that happen elsewhere, such as a new file in a folder, a Telegram message, an MQTT topic, a webhook or a changed web page, and then acts. That always-on loop of waking, planning and using tools is the difference from a chat tool. It is built on the pi coding agent SDK and drives agents from a terminal UI or from Telegram.
Read the full background
Pine brings no model of its own. You sign in to a provider with pine login, for example a Claude or ChatGPT subscription, or point it at a local llama-server, then pick the model in the UI. Agents run shell and file tools inside an operating-system sandbox (Apple's sandbox-exec on macOS, bubblewrap on Linux), and each agent has a mode that decides how much it does without asking. Tools, MCP servers, skills and schedules are set up by asking an agent in plain language, not by editing configuration files.
It fits people who want a home or work automation layer: watching a folder of scanned documents, reacting to a device on an MQTT broker, running coding sessions with a git diff pane, or splitting a job across sub-agents and pinned tabs. Models it can use include Claude and ChatGPT through their subscriptions.
Pine is the wrong pick for Windows users, since the README says only Docker is offered there, and for teams that need vendor support or a long track record. A self-hosted gateway behind your chat apps is OpenClaw. A coding agent for the terminal, desktop and IDE is OpenCode, whose Hokai record cites 208K GitHub stars. A model-agnostic agent with a far larger community is Hermes Agent, whose Hokai record cites past 260,000 stars against Pine's 50 on 12 November 2026. One Docker container per session is NanoClaw. The wider field is in personal assistant agents.
Sources and guides
Official sources
- Pine official site Vendor
- GitHub repository GitHub
Autonomy, guardrails, integrations and pricing on this page are read from pmndrs's own documentation. The verdict is Hokai's reading of them, not the vendor's claim.
Lists that include Pine
More AI Agents on Hokai
NodOpen-source personal agent that works in your apps and browser
Claude DuetAnthropic's proactive agent that briefs you daily and asks before it actsNotisMessage it a task; it works across 1,500+ apps and chases the follow-up
Gemini TaskGoogle's background agent in the Gemini app, formerly called Gemini Spark
TomoA personal AI in your iMessage for goals, habits and accountability.
MuseMeta's AI agent for errands and, since September 2026, small business work.