Open source · Self-hosted · Event-driven

The self-hosted AI agent that reacts the moment things happen.

Cremind is an open-source personal AI agent you run on your own machine. It uses a focused model for every tool call to cut token cost, reacts to real-world events in real time instead of polling, and connects to your Gmail, Calendar, and Jira without you ever setting up a cloud project.

Install Cremind Read the docs View on GitHub

MIT licensed
Runs on your machine
Your data stays local
Works with any LLM provider

What is Cremind

Your own AI agent — not a service you rent

Cremind is an open-source AI agent platform that runs entirely on your own machine or server. It plans and reasons with a powerful model, then carries out work through a unified tool registry that spans built-in tools, Agent Skills, MCP servers, and other agents over A2A.

Install it with a single command, point it at the LLM provider you already pay for — Anthropic, OpenAI, or Groq — and pick how it runs: a one-click Docker desktop, or a native Python install backed by SQLite. Multi-profile isolation keeps work, personal, and experimental setups cleanly apart.

Why Cremind

Five things most agent platforms get wrong

Cremind is built around cost, immediacy, extensibility, and trust — the places where rented, cloud-hosted agents tend to fall short.

Two model groups, one focused agent

Cremind splits work across a high group for reasoning and planning and a low group for individual tool calls. The expensive model thinks; cheap, fast models do the legwork — so you stop paying premium rates for routine tool use.

Real-time, event-driven — never polling

When an email lands or a calendar event changes, Cremind acts within sub-seconds. No cron jobs, no polling loops, no wasteful heartbeat checks burning tokens while nothing happens.

Agent Skills with live event listeners

A Skill is just a SKILL.md plus a few scripts. Skills can ship a long-running listener that pushes real-world events straight into the agent — so the community can build skills that react, not just respond.

Connect: auth you can actually trust

Cremind Connect is a fully open-source broker. The org stores none of your data and never holds your tokens — for Google it isn't even in the token path. Verify it yourself; the whole relay is public.

Connect: OAuth without the homework

No cloud project, no Pub/Sub topic, no OAuth consent screen to register. Click approve in your browser and you're connected — secure, transparent, and private by design.

Cost model

One expensive model for everything? That's the old way.

Most agents route every step — planning and each tool call alike — through a single frontier model. Cremind doesn't.

High group — reasoning & planning

The model that decides what to do. Used for the reasoning loop, where judgment matters and capability is worth paying for.

Low group — per-tool execution

Cheap, fast models that run individual tool calls like browsing or shell commands. Each built-in tool declares its own default group.

Every tool declares its own model group

# tool config (excerpt)
browser              default_model_group = "low"
exec_shell           default_model_group = "low"
markdown_converter   default_model_group = "low"

reasoning loop   ->  high group   (plan & decide)
tool calls       ->  low group    (do the work)

Configured once in the setup wizard and stored in SQLite, with a TOML fallback in settings.toml. Override the group for any tool whenever you want — the result is lower token cost without giving up capability where it counts.

How events flow

From real-world change to agent action — in sub-seconds

Cremind's event system is push-based end to end. Nothing wakes up on a timer; the agent runs the instant something actually happens.

Nudge

A Skill's event listener holds a long-lived WebSocket to the Cremind Connect relay. On an external change, the relay sends a single content-free resync nudge.

Resync & drop

The listener pulls just the delta using its own local token and writes a markdown file into the skill's events/<event_type>/ folder.

Fan out & run

A filesystem watcher detects the file and fans it out to every subscribed conversation. The agent runs immediately, with the event as a synthetic message, streamed back over SSE.

The old way: polling

Check every N minutes. You pay for constant checks that mostly find nothing, and you still react minutes late.

The Cremind way: events

Stay idle at zero cost until something changes — then act in sub-seconds. No cron, no heartbeat, no waste.

How it compares

Cremind vs. typical agent platforms

Comparison of Cremind with typical agent platforms
Capability	Typical agent platform	Cremind
Model usage	One expensive model for planning and every tool call	High group plans; low group runs tools — lower token cost
Reacting to the world	Polling, cron, or heartbeat loops	Push-based events, sub-second latency
Extensibility	Request/response tools only	Agent Skills with long-running event listeners
Integrations (Gmail / Calendar / Jira)	You register your own OAuth app and cloud project	Cremind Connect handles it — click approve in your browser
Your data & tokens	Often flow through a vendor's cloud	Stay on your machine; Connect stores nothing
Hosting	Vendor-hosted SaaS	Self-hosted: Docker or native Python
License	Proprietary	Open source (MIT)

Who it's for

Built for the people who run it and the people who extend it

Get running in minutes

For end users

One install command. A browser-based setup wizard picks your LLM provider, deployment type, and first profile. Connect your accounts with a click — no cloud setup, ever.

Start the quickstart

Open ecosystem

For developers

Write Agent Skills with a SKILL.md and a few scripts. Tap MCP, A2A, and the unified tool registry. Read the architecture, then open a PR.

Build an Agent Skill

Open source

Open source, end to end

Cremind and Cremind Connect are both public and MIT-licensed. The agent runs on your hardware; the relay stores none of your data and never holds your tokens. You can audit every line.

Browse Cremind Browse Cremind Connect

Learn how Cremind Connect keeps authentication private

Built by cremind-ai. Contributions welcome — issues, skills, and pull requests all help.

Documentation

Everything you need, in one place

This site is also the official documentation portal — for end users and open-source contributors alike.

Features

What Cremind does and the concepts behind it.

Architecture & design

How the runtime, tool plane, and events fit together.

Installation

Docker or native — get running in minutes.

Setup & configuration

The setup wizard, LLM providers, and model groups.

Agent Skill development

Build skills with SKILL.md and live event listeners.

Cremind Connect

The token-less OAuth broker and event relay.

Contributing & maintenance

Dev setup, testing, releases, and migrations.

Run your own agent. Keep your own data.

Install Cremind in one command, point it at your LLM provider, and let it react to your world in real time.

# Linux / macOS
curl -fsSL https://cremind.io/install.sh | bash

# Windows (PowerShell)
iwr -useb https://cremind.io/install.ps1 | iex

Read the install guide Star on GitHub