A real-time multi-agent chat interface built with Phoenix LiveView and the Jido agent framework. Create workspaces, add AI agents, and watch them collaborate — with persistent conversations and agent-to-agent communication.
This is a Mix umbrella project containing independently publishable Hex packages and a demo application:
murmur/
├── apps/
│ ├── jido_murmur/ # Core backend — Ingress, Runner, Plugins, Storage, Schemas
│ ├── jido_murmur_web/ # Optional LiveView chat components
│ ├── jido_tasks/ # Task management tools for agents
│ ├── jido_arxiv/ # arXiv academic research tools
│ └── murmur_demo/ # Reference Phoenix app using all packages
├── config/ # Shared umbrella configuration
└── mix.exs # Umbrella root
| Package | Description | Hex |
|---|---|---|
jido_murmur |
Multi-agent orchestration — Ingress, Runner, Plugins, Storage, Schemas |  |
jido_murmur_web |
Pre-built LiveView components for chat UI |  |
jido_tasks |
Task management Jido.Action tools |  |
jido_arxiv |
arXiv search and paper display tools |  |
- Multi-agent workspaces — Add multiple AI agents, each with independent chat history
- Agent-to-agent messaging — Agents communicate via the "tell" tool, with ingress-coordinated busy-run delivery
- Real-time streaming — Token-by-token responses over WebSocket
- Persistent conversations — History survives server restarts via hibernate/thaw
- Autonomous execution — Agents continue processing server-side during disconnects
- Artifacts — Agents produce rich artifacts that affect the UI (e.g. the arXiv agent can display papers)
- Shared task board — Agents manage tasks collaboratively, allowing long-running convergence on complex goals
- Split & unified views — Side-by-side agent columns, or a merged timeline with
@mentionrouting
Prerequisites: Elixir, Erlang/OTP, PostgreSQL (or Docker), and an OPENAI_API_KEY.
# Start PostgreSQL via Docker
docker compose up -d
# Install deps, create DB, run migrations
mix setup
# Start the server
mix phx.serverVisit localhost:4000, create a workspace, add agents, and start chatting.
flowchart TD
subgraph Umbrella
subgraph jido_murmur[jido_murmur — Core]
Ingress --> Runner
Runner --> LLM
StreamingPlugin --> PubSub
ArtifactPlugin --> PubSub
StorageEcto[Storage.Ecto] --> DB[(PostgreSQL)]
end
subgraph jido_murmur_web[jido_murmur_web — UI]
Components[LiveView Components]
end
subgraph jido_tasks[jido_tasks — Tools]
AddTask & UpdateTask & ListTasks
end
subgraph jido_arxiv[jido_arxiv — Tools]
ArxivSearch & DisplayPaper
end
subgraph murmur_demo[murmur_demo — Demo App]
LiveView --> Components
LiveView --> Ingress
end
end
User --> LiveView
PubSub -.->|stream signals| LiveView
When a user sends a message, JidoMurmur.Ingress decides whether to start a fresh run or steer an active one. Tokens stream back in real-time via PubSub. If the agent decides to collaborate, it uses the tell tool to deliver follow-up input through the target agent's ingress coordinator — kicking off a parallel conversation without a Murmur-owned semantic queue.
sequenceDiagram
actor User
participant UI as LiveView
participant RA as Alice
participant RB as Bob
participant PS as PubSub
User->>UI: send message
UI->>RA: ingress deliver
RA-->>PS: stream tokens
PS-->>UI: render live
Note over RA: Alice tells Bob
RA->>RB: tell("Bob", question)
RB-->>PS: stream tokens
PS-->>UI: render Bob live
RA->>PS: done
RB->>PS: done
PS-->>UI: finalize
To use jido_murmur in your own Phoenix application, see the jido_murmur README for installation and configuration instructions.
mix test # Run all tests across the umbrella
mix test --app jido_murmur # Run tests for a specific package
mix precommit # Format + compile + lint + credo + sobelow + test