Features · Build with AI

Describe the job.Approve the plan.

Synapse AI, the assistant inside SynapseOrch, turns a job told in plain words into a workflow. It asks what you didn’t say, shows you a plan you can read, and builds nothing until you approve it.

Free plan, no card needed.

A morning weather check for a café’s patio, built and run in a demo workspace. Played without sound.

Step by step

From a sentence to a running workflow.

  1. Step 1 of 5

    You describe the job

    In plain words, the way you would tell a colleague: what starts it, what it does, where the result goes, and what it should check with you first. Name the site, the API or the app if you know it.
  2. Step 2 of 5

    It asks what you didn’t say

    Before it plans, it reads the sites and APIs you named and looks up what it needs to know. What only you can answer, it asks in one form: here, the forecast address it would use.
    • Reads up to three sites or APIs your message names
    • Answers questions about SynapseOrch from its own docs
    • Asks what only you know, in one form
  3. Step 3 of 5

    You read the plan

    Nothing is built yet. The plan says what will happen, in words you can check:
    1. How it starts. What starts each run: here, its schedule, every morning at 7:00 Lisbon time.
    2. What it will do. One numbered row per step, said in plain words.
    3. Who does each step. A tool, logic or a model, and a person where you sign off. Logic calls no model.
    4. What a run costs. Said before anything runs. This plan calls no model, and a run costs up to about $0.003.
    5. What it assumed. What it filled in that you didn’t say, so you can correct it before it builds.
    6. Missing or off. Anything it can’t do yet, said before you approve. Here: the schedule starts paused.
    7. Your answer. Build it, Change something (it redrafts from what you type), or Stop instead.
  4. Step 4 of 5

    You approve. It builds.

    Press Build it, and it makes the agents and tools the plan needs, saves the workflow and checks it. Then it takes you to what it built, or leaves you the way there if you have moved on.
    • A key goes on a Connect card, never into the chat
    • A schedule starts paused, so the first run is yours to check
    • A workflow that never passes its check is removed again, with the reason
  5. Step 5 of 5

    It runs the way you approved

    Every run follows the saved workflow, and its page says so: the result, the path it took, and how long each step took and what it cost. This one called no model, as its plan said: 0 tokens, $0 model cost.
Synapse AI’s panel: “What can I do for you? … describe a job in plain words, and I’ll show you a plan to approve before I build anything”, and the request under it: every morning, check tomorrow’s forecast for Harborpine Coffee’s patio in Porto and tell me whether to put the tables out; more than a 40% chance of rain, or wind over 30 km/h, means no.
  1. Step 1 of 5

    Synapse AI’s panel: “What can I do for you? … describe a job in plain words, and I’ll show you a plan to approve before I build anything”, and the request under it: every morning, check tomorrow’s forecast for Harborpine Coffee’s patio in Porto and tell me whether to put the tables out; more than a 40% chance of rain, or wind over 30 km/h, means no.

    You describe the job

    In plain words, the way you would tell a colleague: what starts it, what it does, where the result goes, and what it should check with you first. Name the site, the API or the app if you know it.
  2. Step 2 of 5

    Needs your answer: before I show you the plan, check what I filled in that you haven’t told me. “Fetch Porto Forecast” uses this address, which you haven’t given me. Is it the right one? An Open-Meteo forecast address for Porto is filled in, with Stop instead and Continue.

    It asks what you didn’t say

    Before it plans, it reads the sites and APIs you named and looks up what it needs to know. What only you can answer, it asks in one form: here, the forecast address it would use.
    • Reads up to three sites or APIs your message names
    • Answers questions about SynapseOrch from its own docs
    • Asks what only you know, in one form
  3. Step 3 of 5

    The plan: How it starts, on its schedule every morning at 7:00 AM Lisbon time; What it will do, two numbered steps, Fetch Porto Forecast (Tool) and Apply Patio Rules (Transform); no agents, so it runs on tools only with no model called, and a run costs up to about $0.003; What I assumed, the time, Open-Meteo’s free forecast API with no key, and that the result shows on the Runs screen; Missing or off: its schedule is made paused; and Stop instead, Change something and Build it.

    You read the plan

    Nothing is built yet. The plan says what will happen, in words you can check:
    1. How it starts. What starts each run: here, its schedule, every morning at 7:00 Lisbon time.
    2. What it will do. One numbered row per step, said in plain words.
    3. Who does each step. A tool, logic or a model, and a person where you sign off. Logic calls no model.
    4. What a run costs. Said before anything runs. This plan calls no model, and a run costs up to about $0.003.
    5. What it assumed. What it filled in that you didn’t say, so you can correct it before it builds.
    6. Missing or off. Anything it can’t do yet, said before you approve. Here: the schedule starts paused.
    7. Your answer. Build it, Change something (it redrafts from what you type), or Stop instead.
  4. Step 4 of 5

    After approve: Workflow “Harborpine Coffee – Morning Patio Check” created and Schedule “Harborpine Coffee – Morning Patio Check” created, in green, then Synapse AI’s summary of what it built, what each step does, and that the schedule is paused until it is turned on under Schedules, and a card with Open for the new workflow.

    You approve. It builds.

    Press Build it, and it makes the agents and tools the plan needs, saves the workflow and checks it. Then it takes you to what it built, or leaves you the way there if you have moved on.
    • A key goes on a Connect card, never into the chat
    • A schedule starts paused, so the first run is yours to check
    • A workflow that never passes its check is removed again, with the reason
  5. Step 5 of 5

    The run’s page: Completed, 5 seconds, 2 steps, 0 tokens, $0 model cost; the result, “No tables outside tomorrow. Reason: rain chance 88% (above 40%)”; “Followed the saved workflow: 2 of 2 steps, in order”; and the path, each step’s bar on a time scale: Fetch Porto Forecast (Tool, 3.5 s) and Apply Patio Rules (Transform, 1.3 s), neither with a cost.

    It runs the way you approved

    Every run follows the saved workflow, and its page says so: the result, the path it took, and how long each step took and what it cost. This one called no model, as its plan said: 0 tokens, $0 model cost.

Other ways in

Start from what you already have.

An API’s spec, a tool that needs fixing, or code from another tool. Each goes the same way: you see the change before it is made.

From an API

Give it the spec. It makes the tools.

Paste a spec, a docs link, the API’s address or a curl. It makes a tool for each call the job needs, and an agent that uses them. Here it builds an exchange-rate agent from a public API and asks it a question.

Fix what you have

Say what’s wrong. See the diff.

Say what’s wrong with a Python tool. It shows the fix as a diff, with what its test printed, and changes nothing until you press Go ahead. A change to a workflow or an agent is shown field by field, before and after.

Change with AI

Leaving OpenAI Agent Builder? Paste its Agents SDK export, in Python or TypeScript. The code is read, never run, and each agent’s instructions come across word for word; any key written in the code is taken out first. How switching works

Your rules

What it never does without you.

It builds nothing until you approve the plan
Change something as often as you like. Stop instead leaves your workspace as it was.
It shows a change before it makes one
Ask it to change what exists and it answers with the change, each field before and after. Only Go ahead makes it.
It never takes a key in the chat
Keys go on Connect cards, to the screens that keep them. A key pasted anyway is taken out before anything reads it.
It starts schedules paused
You check what a run starts with, then turn it on, or ask it to.
It can undo its last change
Undo is a change of its own, shown to you and approved like any other.
It works within your role
On Cloud a viewer can ask and read; building and changing take the builder role. Its changes are in the audit log.

Nothing hidden

It is built the way you build.

Synapse AI is an ordinary SynapseOrch workflow, with ordinary sub-agents. See how I’m built, in its panel, opens that workflow on a read-only canvas: where it starts, where it stops for you, and the steps in between. Every step type in it is one you can use on your own canvas.

When you run SynapseOrch yourself, Duplicate to my workspace copies it into your workflows to read or change.

Models and cost

What it runs on, and what it costs.

On SynapseOrch Cloud
It builds on Claude Sonnet 4.6 unless your workspace picks its own builder model, paid from your model wallet while it holds money, then on your default model. Every plan includes $2 of models a month, no API key needed. Its chat uses credits, like a run.
On your own servers
It is a new install’s default agent, on the builder model you choose in Settings. Use Claude Haiku 4.5 or a stronger model; reasoning models do noticeably better.
What a build makes
Ordinary workflows, agents and tools: you can open, change, export or delete each one like anything you made by hand.

See pricingThe full guide to Synapse AI

Describe the job your team repeats.

Let us run it, with nothing to install and no API key needed. Or run it on your own servers, free and open source.

Free plan, no card needed.