nextlevelbuilder / goclaw

GoClaw **GoClaw** is a multi-agent AI gateway that connects LLMs to your tools, channels, and data — deployed as a single Go binary with zero runtime dependencies. It orchestrates agent teams and inter-agent delegation across 13+ LLM providers with full multi-tenant isolation. A Go port of OpenClaw with enhanced security, multi-tenant PostgreSQL, and production-grade observability. What Makes It Different • **Agent Teams & Orchestration** — Teams with shared task boards, inter-agent delegation (sync/async), and hybrid agent discovery • **Multi-Tenant PostgreSQL** — Per-user workspaces, per-user context files, encrypted API keys (AES-256-GCM), isolated sessions — the only Claw project with DB-native multi-tenancy • **Single Binary** — ~25 MB static Go binary, no Node.js runtime, 1 GB | 5 s | **GoClaw unique strengths:** Only project with multi-tenant PostgreSQL, agent teams, hooks system, knowledge graph, and MCP protocol support. Architecture Multi-Agent Orchestration GoClaw supports four orchestration patterns for agent collaboration, all managed through explicit permission links. Agent Delegation Agent delegation enables named agents to delegate tasks to other agents — each running with its own identity, tools, LLM provider, and context files. Unlike subagents (anonymous clones of the parent), delegation targets are fully independent agents. | Mode | How it works | Best for | |------|-------------|----------| | **Sync** | Agent A asks Agent B and **waits** for the answer | Quick lookups, fact checks | | **Async** | Agent A asks Agent B and **moves on**. B announces the result later | Long tasks, reports, deep analysis | **Permission Links** — Agents communicate through explicit **agent links** with access control: | Direction | Meaning | |-----------|---------| | | Source can delegate TO target | | | Target can delegate TO source | | | Both agents can delegate to each other | **Concurrency Control** — Two layers prevent any agent from being overwhelmed: | Layer | Config | Example | |-------|--------|---------| | **Per-link** | | support → research: max 3 | | **Per-agent** | | research-bot: max 5 total | **Per-User Restrictions** — The JSONB on agent links supports per-user deny/allow lists. **Agent Discovery** — Each agent has a field for discovery. With ≤15 targets, auto-generated is injected into context. Delegation uses subagent spawning for larger target sets. Delegation vs Subagents | Aspect | Subagents | Agent Delegation | |--------|-----------|-----------------| | Target | Anonymous clone of parent | Named agent with own identity | | Provider/Model | Inherited from parent | Target's own configuration | | Tools | Parent's tools minus deny list | Target's own tool registry + policy | | Context files | Simplified system prompt | Target's own SOUL.md, IDENTITY.md, etc. | | Session | Shared with parent | Isolated (fresh per delegation) | | Permission | Depth-based limits only | Explicit with direction | | User control | None | Per-user deny/allow via settings JSONB | | Concurrency | Global + per-parent limits | Per-link + per-target-agent limits | Agent Teams Teams enable coordinated multi-agent workflows with a shared task board and peer-to-peer messaging. • **Team roles** — Lead agent orchestrates work, member agents execute tasks • **Shared task board** — Create, claim, complete, search tasks with dependencies. Atomic claiming prevents double-assignment • **Team mailbox** — Direct peer-to-peer messaging (send, broadcast, read unread) • **Tools**: for task management, for mailbox Features LLM Providers • **13+ providers** — OpenRouter, Anthropic, OpenAI, Groq, DeepSeek, Gemini, Mistral, xAI, MiniMax, Cohere, Perplexity, DashScope (Qwen), Bailian Coding, and any OpenAI-compatible endpoint • **Anthropic native** — Direct HTTP+SSE integration with prompt caching ( ) for ~90% cost reduction on repeated prefixes. Also supports Claude CLI mode (stdio + MCP bridge with session management) • **OpenAI-compatible** — Automatic prompt caching for OpenAI, MiniMax, OpenRouter (cache metrics tracked in traces). Also supports Codex mode (gpt-5.3-codex via OAuth with "phase" metadata) • **Extended thinking** — Per-provider thinking mode: Anthropic (budget tokens), OpenAI-compat (reasoning effort), DashScope (thinking budget) with streaming support Agent Orchestration • **Agent loop** — Think-act-observe cycle with tool use, session history, and auto-summarization • **Subagents** — Spawn child agents with different models for parallel task execution • **Agent delegation** — Sync/async inter-agent task delegation with permission links, concurrency limits, and per-user restrictions • **Agent teams** — Shared task boards with dependencies, team mailbox, and coordinated multi-agent workflows • **Delegation history** — Queryable audit trail of all inter-agent delegations • **Concurrent execution** — Lane-based scheduler (main/subagent/delegate/cron), adaptive throttle for group chats Tools & Integrations • **60+ built-in tools** — File system, shell exec, web search/fetch, memory, browser automation, TTS, and more • **MCP integration** — Connect external MCP servers via stdio, SSE, or streamable-http with per-agent/per-user grants • **Hooks system** — Event-driven hooks with command evaluators (shell exit code) and agent evaluators (delegate to reviewer) for output validation Messaging Channels • **Telegram** — Full integration with streaming, rich formatting (HTML, tables, code blocks), reactions, media, forum topics (per-topic config and session isolation), speech-to-text, bot commands, group file writer restrictions • **Slack** — Channel integration with bot commands • **Feishu/Lark** — Streaming card updates, media attachments (images/files), mention resolution, topic session mode • **Zalo OA** — Official Account integration for DM conversations • **Zalo Personal** — Unofficial reverse-engineered protocol supporting DM + group messages with restrictive default policies • **D…