Talon splits its runtime into a few clear roles.
Core components
Gateway
The gateway owns the external API surface:
- gRPC service definitions in
proto/gateway.proto - REST endpoints via
google.api.httpannotations - browser-facing UI session endpoints for Sightline and AI SDK-compatible clients
Worker
The worker processes agent execution:
- receives session work
- resolves models and tools
- executes turns
- emits session-step events and persisted session state
The gateway accepts and persists work. The worker executes it.
Sightline UI
The UI is a browser-native operator surface for:
- exploring namespaces and agents
- inspecting sessions and schedules
- observing tool calls and streamed responses
- debugging local or remote Talon deployments
Envoy / edge surface
Envoy provides a consistent edge interface over the gateway. It is the right place for browser-facing concerns like CORS and consistent URL routing.
Resource model
The most important Talon resources are:
- namespaces
- agent templates
- agents
- sessions
- schedules
- namespace knowledge
- MCP servers and bindings
Read Resource Model for the durable control-plane view.
Talon is organized around a few resource types:
- Namespaces: tenancy and grouping
- Agent templates: reusable base definitions
- Agents: instantiated runtime entities
- Sessions: durable conversational/execution state
- Knowledge: namespace or agent knowledge assets
- Schedules: wakeups and recurring work
- MCP servers / bindings: tool connectivity and policy
Transport model
- gRPC is the canonical service contract
- REST is derived from the same proto definitions through HTTP annotations
- Browser streaming is exposed for clients like Sightline and AI SDK-compatible frontends