x-dog

Tools for building AI agents and managing LLM deployments.

GitHub →

x-dog / Packages / claw / Features

What claw does today. Where agent is one loop, claw is the runtime that keeps many agents and their state alive between conversations.

Runtime

Gateway daemon
Unix-socket JSON-line server with streaming, PID/signal handling, and a scheduler loop.
Groups
Isolated per-group runtimes with their own workspace, memory, goals, and tools.
Per-group config
model_id, thinking_level, temperature, max_tokens per group.
Concurrency control
Global semaphore (default 3) + per-group locks; user messages bypass the cap.
Queue modes
collect / steer / steer-backlog with debounce and a backlog cap.

Memory & prompt

Workspace system prompt
IDENTITY / AGENTS / SOUL / USER .md + MEMORY snapshot + one-time BOOTSTRAP.
Long-term memory
Daily logs + MEMORY.md persisted across runs.
Semantic search
Optional vector + BM25 hybrid (RRF), with a keyword fallback when the search extra is absent.
Prompt caching
Static base cached; dynamic workspace overrides applied per turn.

Autonomy

Goals & planning
State-machine + LLM planner with script / agent verification and autonomous delivery.
Scheduling
Cron, interval, and one-off tasks; daily / idle session resets.
Cross-group tools
send_message / goal / task / todo / memory / skill tools between groups.

Channels

tui channel
Local terminal client to a running gateway.
weixin channel
WeChat bot: long-poll monitor, QR auth, typing indicators, context persistence.
Chunked delivery
Paragraph-aware chunking streams long replies naturally.