Sayri will soon be called Sayme. Same project, new name. (And no legal issues)
It can talk, see, listen and do anything both from its multiple UI's and through your favorite messaging app. A thousand times less token spending than Openclaw or Hermes, different sandbox levels and freedom for you.
Sayri is in VERY EARLY ALPHA FOR OTHER OS. we recommend using Pulsar OS for a better experience
Features
Security does not depend on the model, it is programmed. Level 0 never executes anything, levels 1 and 2 run inside bubblewrap, level 3 runs as your user, level 4 elevates through a Polkit dialog you must approve.
A decent harness, that doesn't bombard the API with millions of input tokens. With savings in mind, made by and for those of us who don't have millions to spend on Claude.
whisper.cpp transcribes, Piper speaks, wake word included. The audio never leaves the machine, less token usage and spend; only the LLM API receives your questions.
Run multiple agents in parallel with different sandboxes and gateways without having to read documentation. Intuitive interface, just the way you like it.
Talk to Sayri from Telegram or Discord (and the app you want) without anyone being able to hack you, access your AI or without you misconfiguring it. Create gateways in one click with your favorite AI.
Install from the Pulsar OS store or from ClawHub. OpenClaw-style SKILL.md files work as they are: search once, install from either source, and the agent lists them in its own prompt.
Rules are {action, resource, effect} with whole-value wildcards. System policies can only tighten what your own rules allow, never widen it.
Interfaces
Do you want to look like a millionaire as if you had Siri or do you want to go back to the glory years of Windows? We have a UI for every wish!
A circular indicator pinned to the screen reacts to your audio level and animates through idle, listening, thinking and speaking. The card next to it, the cajita, shows live transcriptions, replies and tool output.
GTK4, layer-shell, WebKit. The default interface.
Clippy, Bonzi Buddy, Merlin, Rover, Peedy, Genie, Links, Rocky and F1 live on the desktop with speech balloons and sound effects. Under the nostalgia they speak the same IPC protocol as everything else.
Back to the 90s/2000
Security
The AI does not decide security, it is decided by the code itself, depending on the sanbox level you configure. Don't trust AI, we protect your agent from prompt injection and malicious vectors
| Level | Name | What it does |
|---|---|---|
| 0 | LEVEL_0_NO_EXEC | Pure conversation. No command ever executes. |
| 1 | LEVEL_1_READONLY | Read-only filesystem, ephemeral /tmp, bubblewrap container. |
| 2 | LEVEL_2_ISOLATED_DEV | Read-only system plus a persistent workspace per agent. |
| 3 | LEVEL_3_HOST_USER | Normal execution as your user. The default. |
| 4 | LEVEL_4_HOST_ROOT | Elevated via pkexec, always with a graphical confirmation. |
The details, including the rule grammar and the policy files: sandbox and security.
We are ready for Linux, perfect in Pulsar OS and preparing for Windows and Android
Grab the current build from the releases page. Questions about Sayri? Discord and Telegram.
Use it without paying for the model
Sayri talks to any OpenAI-compatible endpoint. That includes options that cost nothing, so the agent itself stays free to run.
The innovative ternary and 1-bit models will provide excellent speed to your computer and minimal intelligence.
Get the plugin
Use intelligent AI models for free WITHOUT CREDIT CARD. Daily limits, no hassle.
Open OpenRouter
The most generous, the most intelligent models. Daily limits, worry-free
Open GroqExtend it
Sayri is meant to be widened by the people who use it: teach it a skill, give it a plugin, connect a channel, or draw it a new face.
One markdown file with a name and a description. If it works in OpenClaw, it works here.
Skills guideTalk wherever you want to Sayri and create your own gateway without having to be a development expert
Gateways guideFrom the Siri interface to Microsoft Agents like Clippy or Bonzy Buddy. Yours is next
UIs guidePackage your work for the Pulsar OS store. Submissions are reviewed by AI agents and scanned with VirusTotal.
Store guide
Openclaw was hype but deep down it was slop vibecoded without security awareness or good architecture.
Sayri is the logical evolution, ready to be safe.
Setup, architecture, security, and how to write skills, plugins, gateways and UIs. Generated from the markdown in the repo.
Skills and plugins for Sayri, reviewed by AI agents and scanned with VirusTotal.
Ask about Sayri, share plugins, report what breaks.
The same conversation, on Telegram.
