Architecture &
backend flows.

A technical map of what happens after you press Send.

Message lifecycle

The five layers

PolyphonyAI is built as five layers, from the interface you touch down to where everything is stored. (This summary is shown when JavaScript is off; the interactive version above walks through each layer in turn.)

01 · Interface Layer — the entry point for humans

The native desktop surface. Compose messages, watch agents stream, edit files, and inspect the whole fleet.

02 · Orchestrator — routes, builds context, manages flow

The control center. One message fans out to every agent in parallel, with context assembled and filtered per agent before anything is sent.

03 · Agent Engine — configure, deploy, and run agents

The execution layer. Agents are assembled from a role, a model, and a toolset, then run with permissioned, gated tool access. Code agents can run in their own git worktree, so parallel edits never collide and an excluded agent does not see another's commits until you authorize a merge.

04 · Provider Core — external and local model backends

Direct connections to Anthropic, OpenAI, Google, or a local Ollama. We never proxy your model traffic or credentials through our servers; connections go straight to the provider.

05 · Memory and Storage — long-term memory and system state

Everything persists locally. Append-only memory enforced by the database, semantic recall, git, and the cross-project transfer index.

PolyphonyAI

Built to make complexity visible.