A2A-first Python framework

Build autonomous agents that communicate through A2A-derived tasks.

ProtoLink helps you build distributed, LLM-powered agent systems where every agent is an entity with identity, tools, optional grounded knowledge, transport, discovery and a clean task contract.

Focus on your agent logic, let Protolink handle communication, LLM & Tool integration, authentication, telemetry, logging and storage for you.

Start building Understand the model Read the whitepaper

Try it hereChoose a module, pick an implementation and watch the agent code assemble.

~/protolink/compose1 module

module/transportTransportchoose I/O

composition.pyTransport: HTTP

Documentation paths

Start from the part of the system you are building.

Start

Create an agent

Define an AgentCard, plug in an LLM or tools and start a local or networked agent with minimal boilerplate.

Protocol

Understand the A2A core

Work with AgentCard, Task, Message, Part, Artifact and TaskState as the shared language between agents.

Capabilities

Add LLMs and tools

Use API, server, or local models; expose native Python tools; and adapt MCP tools into the same agent surface.

Systems

Compose flows and meshes

Build coordinators, workers, deterministic flows, registry-backed discovery and multi-agent examples.

Design thesis

Agents are entities, not functions.

An agent can receive work, initiate work, discover peers, expose capabilities, call its model, stream progress and shut down cleanly. The model is one pluggable module inside that entity, alongside tools, transports, state, telemetry, authentication, logging and policy.

Tool-onlyDeterministic capabilities without an LLM.

LLM-onlyReasoning, transformation and conversation.

HybridLLM decisions with typed tool execution.

CoordinatorDiscovery, delegation and multi-agent routing.

What stays stable

A small protocol core, surrounded by swappable modules.

Identity

AgentCard declares name, URL, transport, skills, tags, formats and security schemes.

Work exchange

Tasks carry Messages, Parts, Artifacts, state and metadata across local or remote boundaries.

Capability surface

Native tools, MCP adapters, schemas, examples, LLMs and peer agents become discoverable contracts.

Deployment path

Start in-process with RuntimeTransport, then move the same agent to HTTP, SSE JSON-RPC, WebSocket, or gRPC.

From prototype to mesh

Build one agent, then let the same contracts scale to a system.

The docs cover the full surface: agents, clients, LLMs, tools, RAG, transports, registry discovery, state, storage, telemetry, structured flows, examples and the runtime controls that make production behavior inspectable.

Open the docs index