XDOG

Tools for building AI agents and managing LLM deployments.

GitHub →

xdog / Packages / agent / Features

Reusable SDK facilities for one Agent loop. Applications retain resource ownership, permissions and child scheduling.

Loop

One tool loop
Structured messages and linked tool results; parallel tool calls by default with serial hooks.
Input control
Identified steering/follow-up queues and lifecycle events on the existing Agent.

Lifecycle

Owned shutdown
Cancellation and aclose settle the owned producer; effects are not rolled back or replayed.
Versioned snapshots
dump/restore retain conversation, skill and compaction data; hosts reinject live resources.
Opt-in compaction
One built-in policy; generic Agent compaction is off unless explicitly configured.

Composition

Host-owned tools
ToolRepo supplies actual default bash/filesystem/current_time tools; submit_result is opt-in.
Separate MCP registry
McpRegistry owns MCP connections and proxies; Agent consumes it separately from ordinary tools.
Installed capabilities
CapabilityAssembly combines explicitly opened host bindings and selected skills, without another runtime.
Extension boundary
Workers contribute hooks, commands, UI and providers, not model tools.
Child SDK
Handles and delegation contracts; Coding/Claw own child execution runtimes and authority.