OpenPolly
Open Source · Local-First AI Agent

Your own AI agent. Your own machine.

OpenPolly is a self-hosted assistant with chat, scheduled tasks, memory, and connectors — plus built-in security guardrails that keep your credentials off the LLM entirely. One process, one port, fully open source.

Get StartedView on GitHub
curl -fsSL https://openpolly.ai/install.sh | bash
OpenPolly mascot — a dog holding a wrench
Meet Polly — your always-local helper

AI agents increasingly need real access to your inbox, files, and calendar. Cloud assistants mean trusting a third party with those credentials. OpenPolly runs entirely on your machine — your keys, your data, your rules.

Features

Everything a personal agent needs — nothing it shouldn't have.

Local-first, single process

One process, one port, your machine. No Docker, no accounts, no cloud dependency beyond your chosen LLM provider.

Chat

Streaming responses, tool use, file upload, and an approval flow before the agent takes any risky action.

Tasks

Cron-style scheduled and recurring agent runs — set it up once and let it keep working in the background.

Projects

Persistent, multi-session workspaces that hold file artifacts and notes across every conversation.

Memory

Durable, human-readable markdown notes for long-term memory — no vector database or embeddings required.

Connectors

Gmail, Drive, and Calendar access through your own scoped OAuth client — never a shared third-party app.

Provider flexibility

Bring your own model: OpenRouter, OpenAI, or Anthropic, chosen and verified during onboarding.

Security model

Your credentials never touch the model.

OpenPolly is fully open source and auditable — read the code that guards your credentials.

Agent / LLM
no raw tokens →
Local Token Broker
AES-256-GCM →
Google APIs

A separate local process holds your real OAuth tokens in an encrypted vault. The agent and the LLM it talks to only ever see a scoped capability — never the actual credential.

Attack detection

URL filtering, sensitive-data scanning, and prompt-injection detection, with behavioral risk scoring and rate limiting.

Filesystem sandbox

Opt-in OS-level sandboxing (sandbox-exec, bubblewrap) so agent-run code can't read secrets or write outside scratch dirs.

Inspecting egress proxy

Opt-in outbound traffic routed through an allow/deny proxy with logging — a foundation for future DLP inspection.

Install

Get started in under a minute.

1

Install

One command downloads and installs OpenPolly (via uv) for macOS or Linux.

curl -fsSL https://openpolly.ai/install.sh | bash
2

Onboard

Pick a provider — OpenRouter, OpenAI, or Anthropic — and verify your key.

openpolly onboard
3

Open the app

OpenPolly is now running locally at http://127.0.0.1:7777.

macOS & Linux · MIT licensed · works with OpenRouter, OpenAI, or Anthropic · full installation guide

Ready to run your own agent?

Install OpenPolly in under a minute, or explore the source on GitHub to see exactly how it keeps your credentials safe.

Get StartedStar on GitHub