Purfle

A desktop runtime that installs, schedules, and sandboxes AI agents on your machine.

Purfle runs persistently on Windows and macOS. You install agents — each defined by a signed manifest — and the AIVM (AI Virtual Machine) runs them on a schedule, unattended. Every agent is sandboxed: the runtime enforces what it can access, which LLM engine it uses, and where it writes output. You see one card per agent in the desktop UI. Agents run in the background. The AIVM guards the hen house.

---

Architecture

┌─────────────────────────────────────────────┐
│      .NET MAUI DESKTOP APP                  │
│      (MSIX on Windows / .pkg on macOS)      │
│  ┌──────────────────────────────────────┐   │
│  │              AIVM (C#)               │   │
│  │  Scheduler                           │   │
│  │  ├── AgentRunner: email-monitor      │   │
│  │  ├── AgentRunner: pr-watcher         │   │
│  │  └── AgentRunner: report-builder     │   │
│  └──────────────────────────────────────┘   │
│  UI: one card per agent                     │
└──────────────┬──────────────────────────────┘
               │ talks to
┌──────────────▼──────────────────────────────┐
│           AZURE (live services)              │
│  Key Registry   — Functions (consumption)   │
│  Marketplace    — App Service (F1 free)     │
│  IdentityHub.Web — App Service (F1 free)    │
└─────────────────────────────────────────────┘
               │
┌──────────────▼──────────────────────────────┐
│  purfle demo  — starts these locally        │
│  mcp-file-server :8100                      │
│  mcp-gmail       :8102                      │
│  mcp-github      :8111                      │
└─────────────────────────────────────────────┘

---

Features

Feature	Description
Multi-engine LLM	Gemini, Anthropic, OpenAI, Ollama — engine declared per agent
Signed manifests	JWS with ES256 — every agent is cryptographically signed
5 trigger types	interval, cron, startup, window, event (SSE)
MCP tool protocol	Agents call tools through MCP servers wired by the AIVM
Sandboxed execution	Network, filesystem, env vars — enforced per manifest
Windows + macOS	.NET MAUI — one codebase, two platforms
Marketplace	Publish, search, and install agents from a central registry
Engine-agnostic	No hardcoded engine — the manifest decides
Dashboard	Live summary of all agents with digest, status, and output preview

---

Quick Start

Prerequisites

• .NET 10 SDK
• Node.js (LTS)
• An LLM API key (set GEMINI_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY)

Steps

1. Clone the repository

   git clone https://github.com/clarksonr/purfle.git
   cd purfle

2. Build the runtime and SDK

   dotnet build runtime/src/Purfle.Runtime/Purfle.Runtime.csproj
   cd sdk && npm install && npm run build && cd ..

3. Start the demo environment

   npx purfle demo

   This starts local MCP servers (file, Gmail, GitHub) and prints a summary table.

4. Open the desktop app

   dotnet run --project app/src/Purfle.App/Purfle.App.csproj

   The Dashboard shows installed agents, their status, and today's digest.

---

How Agents Work

An agent is a signed package containing a manifest, optional .NET assemblies, and prompt files:

my-agent.purfle/
├── agent.manifest.json     ← signed, declares everything
├── lib/
│   └── MyAgent.dll         ← .NET assembly (optional)
├── prompts/
│   └── system.md           ← instruction file
└── assets/                 ← optional embedded resources

Minimal manifest (Hello World)

{
  "purfle": "0.1",
  "id": "11111111-1111-4111-a111-111111111111",
  "name": "Hello World",
  "version": "0.1.0",
  "description": "Minimal agent for local demonstration.",
  "identity": {
    "author": "clarksonr",
    "email": "roman@example.com",
    "key_id": "hello-key-001",
    "algorithm": "ES256",
    "issued_at": "2026-03-27T00:00:00Z",
    "expires_at": "2027-03-27T00:00:00Z"
  },
  "capabilities": ["llm.chat"],
  "runtime": {
    "requires": "purfle/0.1",
    "engine": "gemini",
    "model": "gemini-2.0-flash"
  }
}

The AIVM reads this manifest, verifies the signature, checks capabilities, selects the LLM adapter, and runs the agent on its declared schedule. The agent never touches the system directly — the AIVM executes on its behalf.

---

Stack

Layer	Technology
Desktop	.NET MAUI (C#) — Windows + macOS
AIVM	C# class inside MAUI
Manifest spec	JSON Schema Draft 2020-12
Agent identity	JWS / ES256
Inference	Engine-agnostic: Gemini, OpenAI, Ollama, Anthropic
Scheduler	PeriodicTimer + NCrontab + WindowTrigger + EventTrigger (SSE)
SDK / CLI	TypeScript / Node.js
Tests	xUnit + Jest
Azure	Functions consumption + App Service F1
IaC	Bicep

---

Docs

• Getting Started
• Manifest Reference
• Publishing Agents
• Troubleshooting
• Roadmap

---

License

MIT