English | 简体中文

An open-source AI agent platform that does the work for you. Describe any task in plain language — Synapse plans, codes, browses, and delivers results in a secure sandbox, streaming every step in real time.

Chat-driven task execution — Users describe tasks in plain language. Synapse routes each turn into either a single-agent ReAct loop or planner-managed orchestration, then streams progress in real time.

Sandboxed tool execution — Agents can write and run code, install packages, query databases, automate browsers, manipulate files, and generate artifacts inside sandbox sessions created on demand by the configured provider.

Planner-managed multi-agent work — Planner mode can be forced per turn or auto-selected by the execution router. The planner declares a plan, spawns worker agents when decomposition is warranted, waits on results, and streams step and agent status live.

Extensible skill system — Skills are SKILL.md packages with frontmatter for description, allowed tools, dependencies, and sandbox hints. Synapse supports bundled skills plus install flows from git URLs, direct URLs, uploads, and a remote registry.

MCP integration — Connect external tools via the Model Context Protocol. Add MCP servers to extend agent capabilities with third-party APIs and services.

Persistent memory and compaction — Synapse injects user-scoped memory into prompts, compacts long-running conversations with runtime-specific policies, and supports verified fact retrieval for channel-style flows.

Channel integrations — Connect messaging platforms like Telegram to chat with Synapse directly from your favorite apps. Supports bot configuration, account linking, and seamless conversation sync.

Real-time streaming UI — The frontend renders plans, tool calls, thinking, artifacts, browser/computer-use output, and sub-agent progress from the SSE event stream as they happen.

Synapse includes dedicated views for planning, skills, MCP configuration, artifact browsing, and channel integrations.

Multi-Agent Planning	Skills System	MCP Integration

Artifact Library	Channel Integrations

• Theme-adaptive logo: web/public/logo.svg uses strict monochrome lockups via prefers-color-scheme.
• App logo component: web/src/shared/components/Logo.tsx uses monochrome tokens from web/src/app/globals.css (--logo-black, --logo-white, --logo-neutral-700, --logo-neutral-300).
• Favicon variants:
  • SVG: web/public/favicon-light.svg, web/public/favicon-dark.svg (monochrome)
  • PNG/ICO: web/public/favicon-16.png, web/public/favicon-32.png, web/public/favicon.ico, plus dark PNG variants (favicon-dark-16.png, favicon-dark-32.png)
  • Apple touch icons: web/public/apple-touch-icon.png, web/public/apple-touch-icon-dark.png
  • PWA icons: web/public/icon-192.png, web/public/icon-512.png
• Metadata wiring: web/src/app/layout.tsx serves light/dark favicon SVG variants using Next.js icons entries with media.
• Monochrome usage rules: docs/logo-monochrome-spec.md

• Google OAuth authentication with per-user skills and MCP server configurations
• Conversational interface with file upload, skill selection, and follow-up messages
• 20+ built-in tools — web search, code execution, browser automation (with step tracking), computer use (with action metadata), file operations, database queries, image generation, document generation
• Execution routing — use_planner can force planner mode, otherwise the backend classifies each turn into single-agent or planner-managed execution shapes
• Plan mode — Explicit task decomposition with step names, execution types, and live checklist/progress tracking
• Artifact management — Files generated in the sandbox are extracted and available for download/preview, with a dedicated library page for browsing all artifacts
• Extended thinking — Configurable thinking budget for deeper reasoning on complex tasks
• Persistent memory — User-scoped key-value memory, channel facts, and runtime context compaction for long conversations
• Conversation history — Full persistence with PostgreSQL
• Agent evaluation system — YAML-defined eval cases with programmatic and LLM-as-judge grading, covering tool use, skill invocation, sub-agent spawning, and agent handoff
• Channel integrations — Connect messaging platforms (Telegram) to chat with Synapse from your favorite apps
• User preferences — Persistent theme (dark/light/system) and locale settings per user
• Dark/light theme with internationalization (English, Chinese)
• Keyboard-first UX — Command palette (Cmd+K), responsive layout

• Python 3.12+, Node.js (with npm), uv
• PostgreSQL (optional, for conversation persistence)
• Rust 1.77+ (optional, for desktop app)

make install

# Create backend/.env (see backend/.env.example)
# ANTHROPIC_API_KEY=...
# TAVILY_API_KEY=...

make dev

Open http://localhost:3000.

Synapse also ships as a native desktop app built with Tauri v2. It wraps the same web UI in a native window with automatic backend/frontend process management.

# Dev mode
make desktop

# Production build (.app / .msi / .deb)
make build-desktop

The desktop app manages backend and frontend as sidecar processes — if they're already running (e.g. via make dev), it connects to the existing services. Google OAuth opens in the system browser and hands the session back to the desktop window automatically.

See Desktop App Guide for configuration and troubleshooting.

• Local Setup Guide — Step-by-step instructions to get Synapse running on your machine
• Development Guide — Commands, architecture, API reference, environment variables, and contribution workflow
• Documentation Index — Reference shards and deeper docs, including agent runtime, memory, and evals
• Desktop App Guide — Tauri desktop app setup, configuration, OAuth flow, and troubleshooting
• Agent Memory Guide — Working context, compaction, persistent memory, and verified facts
• Agent Implementation Review — Current audit findings and confirmed documentation drift
• Design Style Guide — UI component patterns, color system, typography, and accessibility
• Brand Guidelines — Brand identity, color palette, and visual design language

Apache-2.0