zai-org/Synapse

What it solves

Synapse provides a self-hosted AI workspace designed to move beyond isolated chatbots by creating "digital teams." It solves the problem of fragmented AI interactions by treating the conversation as the primary boundary where humans, native AI actors, and remote agents can collaborate in a single thread with shared memory and governed access to tools.

How it works

Synapse uses a conversation-centric architecture where every interaction is scoped to a specific thread. It integrates several key components:

• Participants: Supports four types of participants, including workspace members, native actors, and bridged remote agents (like Claude Code) that join via a WebSocket daemon.
• Runtimes & Resources: Provides governed execution surfaces such as paired physical devices (Linux servers, desktops) or isolated sandboxes (Docker, E2B) for running code and using tools.
• Governance: A central workspace_resource_grants ledger manages explicit, revocable permissions for plugins, memory spaces, and device capabilities.
• Connectivity: Bridges external IM platforms (Telegram, WhatsApp, etc.) and event sources (GitHub webhooks, schedules) directly into the conversation runtime.
• Memory: Implements permissioned memory spaces using a combination of lexical search and embeddings to maintain context across different conversations.

Who it’s for

It is designed for users and teams who want to host their own AI orchestration layer that manages multiple agents, integrates with their existing hardware/software tools, and maintains a strict security and permission model for AI-driven automation.

Highlights

• Multi-Agent Collaboration: Actors can wake each other up and hand off tasks within the same transcript.
• Remote Agent Bridging: Allows existing external agents to join as first-class participants without migrating their runtime.
• Unified Authorization: Sensitive tool calls can be approved via interactive cards directly in the chat.
• Device Integration: Pairs with real machines to expose filesystem, CLI, and browser capabilities as MCP tools.
• Event-Driven Automation: Schedules and webhooks trigger actor wakeups as visible conversation events.