Open Source — Apache 2.0

The portable AI agent format

.fylle

An open protocol to package, share, and deploy AI agents across any platform. Build once, run everywhere.

Agents are portable. Intelligence is not.

Terminal — fylle-format
# Package your agent
$ fylle pack ./my-agent --output content-curator.fylle
✓ manifest.yaml validated (v0.2)
✓ 2 tools: web_search (mcp), save_notes (mcp)
✓ 1 skill, 2 inputs, guardrails OK
→ content-curator.fylle (3.8 KB)

# Share it
$ fylle push content-curator.fylle
→ Published to hub.fylle.ai/agents/content-curator@1.0.0

# Pull and run on any runtime
$ fylle pull content-curator --runtime langchain
→ Agent ready. Mapped 2/2 tools. Inputs: brand_context, topic.

One file. Complete agent.

A .fylle package contains everything needed to understand, transport, and run an AI agent — without vendor lock-in.

content-curator.fylle
├── manifest.yaml — identity, version, model, inputs, output, tags
├── agent.md — system prompt, tone, role, language
├── skills/ — modular capabilities (prompt-based)
│ └── research.yaml
├── memory-schema.yaml — what to remember (not the data itself)
├── guardrails.yaml — rules, limits, max autonomy
└── README.md — human-readable documentation
📝

Declarative-first

Agents are described in YAML and Markdown, not code. Anyone can read, edit, and understand what an agent does.

🔌

MCP-native tools

Tools are declared with protocol: "mcp" — the runtime maps them to available servers. Graceful degradation for optional tools.

🧩

Composable inputs

Agents declare their inputs explicitly. The runtime knows what to inject — making agents composable in pipelines and multi-agent workflows.

Runtime-agnostic

Works on LangChain, CrewAI, OpenClaw, or any platform that implements a simple adapter. No lock-in, ever.

🧠

Memory-portable

The memory schema travels with the agent, the data doesn't. The agent knows what to remember — the runtime decides how.

🔐

Secrets-safe

API keys and credentials are never stored in the package. Sensitive values are declared as requirements, configured at runtime.

The missing layer.

MCP connects agents to tools. A2A connects agents to agents. But nothing packages and transports the agent itself. Until now.

Marketplace
.fylle Hub
Fylle (open source)
Agent Package
.fylle format
Fylle (open source)
Agent ↔ Agent
A2A Protocol
Google / Linux Foundation
Agent ↔ Tool
MCP
Anthropic
Runtime
LangChain, CrewAI, OpenClaw, Fylle
Various
LLM
Claude, GPT, Gemini, DeepSeek
Various
Standard What it does What it doesn't
MCPConnects agents to external tools & dataDoesn't package or transport agents
A2AEnables agent-to-agent communicationDoesn't define or export agents
Agent File (.af)Exports/imports agents in LettaTied to one ecosystem
ADLDeclares agent definitionsNo memory, skills, or runtime
.fylleFull portable agent — tools, skills, memory, inputsOpen, universal, runtime-agnostic

Who benefits.

The .fylle format is designed to create value across the entire AI agent ecosystem — not just inside one platform.

Indie Developers

Build once, sell anywhere

Package your agent, publish it to the Hub, sell or share it. You're never locked into a single framework or marketplace.

Companies

Standardize your agents

Define how your team builds agents internally. Swap runtimes without rewriting. Audit and version every agent like code.

Frameworks

Import/export as a feature

LangChain, CrewAI, or any runtime can offer .fylle import/export. Lower the barrier to entry. Attract more users.

Marketplaces

A standard listing format

Like npm for packages or Docker Hub for images. A universal format to list, discover, and distribute AI agents.

Open-source the protocol.
Monetize the platform.

The Protocol

Free. Open. For everyone.

The .fylle format is Apache 2.0. Use it to package, share, and deploy agents across any platform. No fees, no lock-in, no strings attached. Agent portability is a public good.

The Platform

Where agents compound.

Fylle the product is the execution layer where agents accumulate permanent context and improve through feedback loops. Every interaction compounds. Your agent on day 100 is fundamentally smarter than on day 1. Agents are portable — the intelligence they build is not.

Build your agent anywhere.
Run it everywhere.
When you want agents that compound
run them on Fylle.

Go deeper.

The protocol is the foundation. These tools bring it to life.

Start building.

The spec is open. The CLI works. The first agents are live.
Join the community shaping how AI agents move.