Introduction
AI automation without the surprises. Describe a recurring job in plain words, and get a workflow you can read, that runs the same way every time and checks with you before anything goes out.
SynapseOrch turns the jobs your team repeats every week into workflows that run themselves — the report, the triage, the brief, the review. You describe the job in plain words, it builds the workflow, and a person signs off wherever the output leaves the building.
Most agent tools let the model decide its own next step, which is why they work ninety-nine times and surprise you on the hundredth. SynapseOrch fixes the path instead: every run follows the steps you approved, on the model you chose, and you can see what each step did and what it cost.
Start on Cloud
Nothing to install. Managed models on a prepaid wallet, and a free plan to start.
Cloud quick start
From a new workspace to your first run.
Self-host it
The same open-source engine on your laptop, your servers or your own cloud.
Build with Synapse AI
Describe the job, and review the workflow it designs for you.
How it works
Describe it
Tell Synapse AI the job in plain words, or start from a template. It draws the workflow: the steps, the model for each, and where a person signs off.
Approve it
Read the workflow on the canvas before it runs. Change any prompt, tool or model, and put a human step wherever the output leaves the building. Tool policy decides which tool calls need a yes first.
Rely on it
Run it on a schedule, from a chat message or through the API. Every run takes the same path, and the logs show what each step did and what it cost.
What you can count on
Same path, every run
A workflow is a fixed path of steps, and execution follows it — no detours the model invents on the way. Where the next action is already known, Tool and LLM steps skip the agent loop entirely: no reasoning overhead, the same shape of output every time, and far cheaper. When a step fails, the run says so instead of carrying on as if it worked.
Nothing goes out without a yes
Human steps pause a run and wait — for minutes or for days — then resume exactly where they stopped when someone answers, in the app or from a chat app. Runs survive restarts, and on Cloud, time spent waiting for a person never costs a credit.
The bill stops where you set it
Every step names its model. A small, fast one routes and classifies; a strong one handles only the steps that need judgement; and some steps need no model at all. On Cloud, a plan stops at its allowance, and nothing is billed past it; managed models draw from a prepaid wallet.
Yours to keep
The engine is open source under AGPL-3.0, and Cloud runs the same one. Your workflows can always move to your own servers, with no licence check and nothing sent to us — nobody can switch them off. OpenTelemetry export to your own collector is there if you want it, and off by default.
Any model, switch any time
Fourteen-plus hosted providers, local models through Ollama, and CLI mode for Claude, Gemini and Codex with no API key at all. Switch an agent's model in the middle of a conversation and it keeps its prompt, memory, tools and plan. On Cloud, managed models need no key.
Turn anything into a tool
- Python → dropped in, it runs sandboxed
- Any REST API → describe the schema, or import an OpenAPI spec
- Any MCP server → local subprocess or remote HTTP
- Any workflow → promote it to an agent and nest it inside another
What people build
Inbox triage
Classify each message, escalate the urgent ones and draft the reply — you approve it before it goes.
Reports and briefs
The weekly report, the market brief, the competitor watch: the same shape every time, reviewed before it is filed.
Scheduled automation
Runs every morning against live data, with the result posted where your team reads it.
Customer-facing chat
An agent wired to your docs, database and APIs, embedded in your own site with one script tag.
AI inside your product
Call a workflow over the REST API and stream structured results back.
Start from a template
Working workflows on Cloud, ranked to match what your business does.
Two ways to run it
| SynapseOrch Cloud | Self-hosted | |
|---|---|---|
| Infrastructure | None — we run it | Yours |
| Models | Platform models on a wallet, or bring your own keys | Your own keys, or local models |
| Team | Workspaces, roles and seats | Single workspace |
| Upgrades | Continuous | You choose when |
| Licence | Commercial | AGPL-3.0 |
Most people should start on Cloud. If you already know you need to run it yourself, Cloud vs self-host walks through the trade-offs.
Key concepts
| Concept | Description |
|---|---|
| Agent | An AI assistant with a system prompt, a set of tools and an optional model of its own. It works in a loop — think, call a tool, read the result — until the task is done. |
| Workflow | A workflow: a fixed path of steps (a DAG). Each step is an agent call, an LLM call, a tool call, a branch, a loop or a human gate. |
| Tool | Something an agent can call — built in, an MCP server, or your own HTTP or Python. |
| Skill | A reusable instruction package an agent loads when a task needs it, rather than carrying it in every prompt. |
| MCP server | A local or remote process exposing tools over the Model Context Protocol. |
| Vault | File storage agents read across sessions, and reference in prompts as @[path]. |
| Session | Conversation context scoped to a session_id. |
| Schedule | An interval or cron trigger that runs an agent or workflow. |
Capabilities
| Capability | Detail |
|---|---|
| Synapse AI | Describe what you want; it designs and builds the workflow |
| Team agents | A lead hands work to its members — one at a time or as a team working in parallel — and answers with what they found |
| Multi-model workflows | A model per step — mix hosted, local and CLI models in one run |
| Workflows as agents | Promote any workflow to an agent; nest pipelines inside pipelines |
| Deterministic steps | Call a tool directly with state values, skipping the agent loop |
| Knowledge bases | Agents answer from your documents, with sources — hybrid search, citations, evaluation |
| Skills and memory | Instruction packages loaded on demand; durable memory across sessions |
| Evals | An LLM judge scores runs against a rubric and catches regressions |
| Tool policy | Decide which tools need a person's approval, per agent or workspace |
| Resumable runs | A per-run journal replays a stream you disconnected from and resumes mid-node |
| Sandboxed Python | Agents write and run Python in an isolated container |
| Embeddable widget | Put an agent in your own site with one script tag |
| 5 messaging platforms | Slack, Discord, Telegram, Teams, WhatsApp — with per-agent channel binding |
| 14+ model providers | Hosted, local and CLI |
| Import / export | Portable bundles of agents, workflows, servers and tools, plus 5 example packs |
Next
Cloud vs self-host
Pick a path.
Guides
Build real things step by step, on one example workspace.
Agents
System prompts, tools and a model of their own.
Workflows
Multi-step workflows on a fixed path.
API reference
Drive it from your own application.
The engine is open source: github.com/synapseorch-ai/synapse-ai, AGPL-3.0.