Multi-agent AI development needs shared state. We build the coordination and memory layer.

When multiple AI agents work on the same codebase:

• 🔄 No shared context — agents duplicate work, one doesn't know what another did
• ⚠️ Conflicting changes — overlapping modifications cause merge conflicts
• 📝 No learning — agents can't learn from past failures across sessions
• 🔍 Grep is dumb — string search finds false positives; agents need graph queries

Synapses OS is three specialized services working together:

┌─────────────────────────────────────────────────────────┐
│              Your IDE (Claude Code, Cursor, Zed, etc.)   │
└────────────────────┬────────────────────────────────────┘
                     │ MCP Protocol (stdio)
                     ▼
┌─────────────────────────────────────────────────────────┐
│  synapses (core)                                         │
│  ├─ Graph engine (nodes: functions, structs, files)    │
│  ├─ 38 MCP tools (context, graph queries, tasks)       │
│  ├─ Episodic memory (past decisions, failures)         │
│  ├─ Agent message bus (broadcast work status)          │
│  └─ SQLite persistence (multi-session state)           │
└────────┬────────────────────────────────────┬──────────┘
         │ HTTP                               │ HTTP
         ▼                                     ▼
    ┌──────────────────┐            ┌──────────────────┐
    │ synapses-        │            │ synapses-        │
    │ intelligence     │            │ scout            │
    │ ├─ 4-tier LLMs  │            │ ├─ Web search   │
    │ ├─ Summaries    │            │ ├─ URL fetch    │
    │ ├─ Insight gen  │            │ ├─ Distillation │
    │ └─ Embeddings   │            │ └─ YouTube      │
    └──────────────────┘            └──────────────────┘
         llama-server                  DuckDuckGo/Tavily
         (local, no Ollama)            (optional API key)

✅ Structured Code Understanding — Graph of functions, structs, calls, dependencies (not grep) ✅ Multi-Agent Coordination — Message bus + work claims prevent conflicting changes ✅ Episodic Memory — Agents learn: "I tried this pattern last week and it failed" ✅ Local LLMs — Brain sidecar (Qwen/Mistral/etc) on llama-server, no cloud ✅ No Code Leaves Your Machine — All inference + search local-first ✅ 18 Languages — Go, TypeScript, Python, Java, Rust, C++, and more ✅ AI-Ready — Design specifically for agent workflows (not human IDE features)

curl -fsSL https://raw.githubusercontent.com/SynapsesOS/synapses/main/install.sh | sh

cd /path/to/your/codebase
synapses init       # Parse code, build graph

synapses start -path /path/to/your/codebase
# Ready for IDE connection

brain setup --llama-server  # Download llama-server + model
brain serve                  # Starts on :11435

# Configure for Claude Code, Cursor, Zed, etc.
synapses mcp-setup --agent <cursor|claude|zed|windsurf>

Then agents can use 38+ MCP tools: get_context, find_entity, recall, remember, send_message, and more.

Session 1 (Frontend Agent):
  agent_id = "frontend-claude"
  session_init()
  → returns: "backend-claude modified /api/auth.go 2h ago"

  Ah, I need to update TypeScript types to match the new auth API
  remember(decision="Updated LoginRequest shape to include mfa_enabled",
           outcome="success")

Session 2 (Backend Agent):
  agent_id = "backend-claude"
  get_messages(agent_id="backend-claude")
  → returns: frontend's memory + recent changes

  Backend can now safely modify /api/checkout.go knowing frontend is aware

With Synapses:

• ✅ No duplicate work (agents see what each other did)
• ✅ No surprises (changes propagate as messages)
• ✅ Learning loop (failures recorded, future agents learn from them)

🚫 No Cloud — All processing local (CPU/GPU) 🔄 Fail-Silent — Brain down? Graph queries still work. No panics. 📦 Single Binary — One MCP server, works with any IDE 💾 SQLite-Only — Pure Go, no C dependencies at runtime 🔁 Incremental — File watcher re-parses only changed files 🧠 AI-Native — Designed for agents, not humans

All your code stays on your machine.

• Graph indexed locally in ~/.cache/synapses/
• LLM inference on localhost (llama-server, no Ollama required)
• Optional web search uses DuckDuckGo (or your Tavily key) only when you call web_search
• No telemetry, no tracking, no cloud uploads

• synapses README — Full feature overview + all MCP tools
• synapses-intelligence README — Brain setup (CPU/GPU auto-detect)
• synapses-scout README — Web search + distillation

• Issues: Report bugs per-repository (synapses / intelligence / scout)
• Discussions: Ask questions on GitHub Discussions
• Security: Report privately to security@synapsesos.dev

We welcome pull requests and issues! Each project has contribution guidelines:

• Go projects (synapses, synapses-intelligence): make build && make test
• Python project (synapses-scout): pip install -e ".[dev]" && make test

All three projects are MIT Licensed.

Made with ❤️ for multi-agent AI development.

📖 Read the Vision · 🚀 Get Started · 💬 Discuss