Multiplayer AI infrastructure

AI is becoming a team sport.
Give it a room.

Oryxa is the shared layer where people and specialised AI agents can work on the same question, live.

What Oryxa enables

Build internal AI applications
with a shared point of view.

Ship customer-facing products
where agents collaborate.

Teams own the application. Oryxa provides the framework and infrastructure underneath it.

One binaryRoom, runtime, viewer and CLI
Seven integrationsHTTP, ACP and major agent frameworks
Apache 2.0Open, inspectable and self-hosted
The thesis

Every agent has a job.
Oryxa gives them a room.

Your prompts, tools, memory and orchestration stay where they are. Oryxa supplies the shared session, the event log and the coordination layer around them.

People

Watch live work, steer a conversation and join from the same shared context.

Agents

Work in parallel lanes, with their own task, cursor and conversation state.

One room

A durable record of what happened, what changed and what everyone needs next.

Many specialists. One shared session. A durable source of truth.

What is underneath

The agent is not the product. Coordination is.

One framework cannot own a conversation shared by several people and several independent systems. That is the layer Oryxa builds.

Shared sessionsEveryone sees the same work, live and replayable.
Parallel lanesAgents answer together without blocking each other.
Shared contextState travels with the room, not a single model call.
Durable eventsThe log is the source of truth after every restart.
The product

One simple surface.
Many different agents.

Connect an HTTP agent or an operator-controlled local coding agent. Oryxa turns both into dependable participants in the same room.

Selective attention

Only agents with a reason to answer are woken. Human conversation stays human.

Independent frameworks

Connectors describe HTTP calls. Your agents keep their prompts, tools and memory.

Live, durable rooms

Every turn streams live, then becomes part of the event log the room can replay.

Operator controls

Rooms, keys and budgets let an application decide who participates and what a turn can cost.

Connecting agents

A connector is a description,
not another codebase.

A small YAML file describes what to call and where the answer lives. The agent does not need to know that Oryxa exists.

connectors/my-agent.yaml
name: my-agent
base: http://localhost:5000

turn:
  method: POST
  path: /run
  body:
    prompt: "{{input}}"
  response:
    format: json
    text:
      - $.output  # first matching path wins
The core

Oryxa

The session runtime owns input, presence, streaming output and the shared event log.

Connected to it

Your agent

  • LangGraph
  • Google ADK
  • Pydantic AI
  • CrewAI
Also supported

Your coding agent

  • Claude Code
  • Codex
  • Any ACP agent
  • Any HTTP service
Start in two minutes

One binary.
Your agents.

One command installs Oryxa and starts it — the room view opens in your browser. Bring your own framework when you are ready.

curl -fsSL https://oryxa.in/install.sh | sh
01
Install and start

One command. It installs the binary, starts the local runtime and opens the room view in your browser.

02
Open a room

Claude Code, Codex and OpenCode are built in and offered when their executable is on your PATH.

03
Invite the team

People join through the embedded browser viewer and follow every turn live.

Open source · Apache 2.0 · Rust

The infrastructure behind
multiplayer AI.