Synapse AI
Synapse AI is a production-grade, open-source multi-agent orchestration platform for building, connecting, and executing AI agents powered by any LLM — local or cloud. It converts existing APIs and Python programs into agent tools, orchestrates them into deterministic end-to-end workflows, and exposes everything via a REST API — without vendor lock-in.
What can you build with Synapse?
- Agentic pipelines — Research → summarise → send report → await human approval → publish
- Customer-facing AI — Chat assistants connected to your database, docs, and internal APIs
- Scheduled automation — Cron jobs that run agents against live data every morning
- Internal tooling — Agents that query your code repos, databases, Jira, and Slack in one shot
- API-first AI — POST to
/api/v1/chatand stream back structured responses
Architecture
Key concepts
| Concept | Description |
|---|---|
| Agent | An AI assistant with a system prompt, tool subset, and optional model override. Runs in a ReAct loop. |
| Orchestration | A DAG (directed acyclic graph) of steps. Each step can be an agent call, LLM call, tool call, branch, loop, or human gate. |
| Tool | A function an agent can call — native (Sandbox, Vault, SQL…), MCP server, or custom REST/Python. |
| MCP Server | A local or remote process exposing tools via the Model Context Protocol. |
| Vault | Persistent file storage accessible to agents across sessions. |
| Session | Conversation context scoped to a session_id. Agents remember prior turns. |
| Schedule | Interval or cron trigger that runs an agent or orchestration automatically. |
What makes Synapse different
Cut costs without cutting quality
Run a different model at every step. Use a fast, cheap model for routing and classification; switch to a powerful model only for the steps that need it. One orchestration, many models — you control exactly where the compute goes.
Workflows that do exactly what you designed
Orchestrations are strict DAGs. Execution follows the exact path you defined — no surprises, no hallucinated detours. For steps where the next action is already known (fetch this, parse that, write here), use Tool and LLM steps instead of full agents: zero reasoning overhead, deterministic output, and far cheaper to run.
Turn anything into a tool
Your existing systems are already the capability — Synapse just makes them available to agents:
- Any Python program → drop it in, it becomes a sandboxed agent tool
- Any REST API or webhook → describe its schema, agents call it natively
- Any MCP server → local subprocess or remote HTTP, connected in seconds
- Any orchestration → promote it to an agent; chain orchestrations like functions
Never blocked on a human decision
Human steps pause execution mid-workflow and wait. When the person responds — via the UI, Slack, Telegram, or any connected messaging channel — the run resumes exactly where it left off. No polling, no timeouts you didn't set.
Run it anywhere, own your data
Full local operation with Ollama. Or mix: local models for some agents, cloud APIs for others. No vendor lock-in on models, no data sent anywhere you didn't choose.
Capabilities at a glance
| Capability | Detail |
|---|---|
| Multi-model orchestrations | Per-step model override — mix Gemini Flash, Claude Opus, and local Ollama in one workflow |
| Orchestrations as agents | Promote any orchestration to an agent; nest pipelines inside pipelines |
| Deterministic Tool steps | Skip the ReAct loop entirely — call a specific tool directly with state values |
| Resumable human gates | Human steps survive server restarts; runs pick up exactly where they paused |
| Docker-sandboxed Python | Agents write and execute Python in an isolated container — safe by default |
| Cost limits per run | Set a max-spend per orchestration run — execution halts if the budget is hit |
| 5 messaging platforms | Slack, Discord, Telegram, Teams, WhatsApp — with per-channel agent binding |
| 14+ LLM providers | Cloud, local, and CLI providers; no API key needed for Claude/Gemini/Codex CLI |
| Import/Export packs | Portable bundles of agents + orchestrations + MCP configs; 3 curated starter packs |
| AI Builder | Chat with a meta-agent that designs and materializes orchestrations for you |
Quick links
- Install Synapse — one-command setup for macOS, Linux, Windows, Docker
- Quick Start — create your first agent and chat in 5 minutes
- Building Agents — configure system prompts, tools, and model overrides
- Orchestrations — build multi-step AI pipelines
- REST API Reference — integrate Synapse into your application
- GitHub →