Synapse AI

Synapse AI is the assistant built into SynapseOrch. Ask it how the product does something, or what ran, what failed or what waits on you. Ask it to run, stop or answer something, and it puts up a card you press. Ask it to change something that exists, and it shows the change for you to approve. Describe a job in plain words, and it researches what it needs, drafts a plan in plain words, then builds the workflow once you approve it. Nothing changes in your workspace until you press a card or approve, and a key never goes into the chat: it goes on a Connect card.

Where to find it

• The menu: the Synapse AI row, on every page. On a phone, the mark in the top bar.
• Jump to (⌘K): Ask Synapse AI.
• Home: the ask line above Needs you, with up to four suggestions under Things to ask: What needs me? while something waits, Why did "Invoice triage" fail? after a failure in the last day, What's running right now?, Summarise today, and What ran overnight? in the morning or What did my agents do today? later. Getting started has Build with AI, and the Coming up tip Schedule it with AI.
• Workflows: Build with AI in the page's header. On the canvas, Change with AI opens it on the workflow you are looking at.
• Chat: on SynapseOrch Cloud, Synapse AI is the first agent in the picker and a new conversation's default. Self-hosted, a new install's default agent is Synapse AI.

The panel and Chat show the same conversation, so a question you started in one is there in the other. The panel's header has See how I'm built and Start a new conversation.

The Agents page's Build with AI is a different tool: it writes one agent's instructions, picks its tools and opens the agent there. See Creating agents.

Asking about your workspace

Synapse AI reads what your role can see: workflows and how their runs went, one run step by step ("did it finish", "why did it fail"), what an agent has been doing, a summary of today or a week, what waits on you, boards and their cards, schedules, channels, tool policy, knowledge bases, usage, your plan and your models. It never shows a secret's value. A question that only reads changes nothing and needs no card.

Cards you press

When you ask Synapse AI to do something the app does with a button, it puts up a card that does exactly that, marked Needs your go-ahead and Nothing happens until you press.

• Stopping and deleting are red. A delete says This cannot be undone.
• A card works once, for ten minutes, and only for you. Pressed again: Already done: this card was pressed before. Too late: This card has expired. Ask again for a new one. Someone else's: This card is not yours to press, or it was changed.
• A card it can't make is said instead: "Monday digest" is already off., or, when two match, the names and ids to choose from.
• After you press, the card says what happened (Started., Stopped., Allowed. The agent carries on.) with a link to the run.

Members, billing, API keys and provider keys stay on their screens. A key is never typed into the conversation: Synapse AI puts up a card that takes you to the screen where it is stored.

Changing what exists

Ask in plain words, for example:

• rename an agent, a workflow or a schedule; move a schedule to another time;
• change an agent's model, tools, skills, instructions or tool policy;
• change a workflow's limits;
• delete an agent, a workflow, a schedule or a tool;
• "undo that".

Synapse AI answers with The change: under What it would do, one numbered row per change, such as Rename the agent "Researcher" — to "Market researcher". Where a change sets fields, its row lists each one as it is now and as it will be, before → after: a field with nothing before is new, and one with nothing after goes, shown (none). When Synapse AI's own words for a change say more than the row, they follow it as Described as: …, so you approve what will be done, not only how it was put. A delete is a red row, Deletes: the agent "Researcher", saying what uses it, or Nothing uses it, and It can come back only as a new copy. A workflow or an agent that something still uses is not deleted: Synapse AI says what uses it, as the app does.

Press Go ahead to make the change, Change something to say what should be different (it redrafts), or Stop instead. Typing something else while the change waits sets it aside: nothing in it was done.

• Two with one name: More than one agent matches "Researcher". Which one do you mean?, with a numbered list of each one's name and id. Answer in the message box.
• Keys never go in a change: a key typed into a value is refused, with where to add it instead. See Keys and sign-ins.
• One request at a time: several separate requests in one message are answered Ask me for them one at a time. After five rounds of changes without settling, it stops and asks you to start again in one message.
• A change to a workflow's steps goes the way a build does: Synapse AI drafts a plan and shows The plan (see below). Making a new agent, skill, HTTP tool or OpenAPI import is reviewed as Make the agent "…", and made once you approve. HTTP tools and Python tool fixes have their own sections.

After you approve, Synapse AI says what changed — Changed the workflow "Invoice triage"., Deleted the agent "Researcher". — and anything that did not take effect, plainly:

• Not done: the agent "Researcher" changed since you approved this (model), so I left it as it is now. Ask again to change it.
• Not done as asked: after the change, the agent "Researcher"'s model should read claude-sonnet-5 but reads …

Undo

Ask "undo that", or name the change. Undo is a change of its own: it is shown on its own card and needs its own Go ahead. Only your own last change can be undone.

• A deleted item comes back as a new copy, with a new id: an agent without its memory, a workflow without its run history, a schedule paused, an HTTP tool without its headers (you enter its key again on its card), and a Python tool without the variables it read (you add them again in its editor).
• A value that held a key was not kept, so it can't be put back, and the card says so.
• A name another item now has can't be taken back, and a change Synapse AI can't reverse is said: That change can't be undone from here.

Building a workflow

Describe the job in plain words: what starts it, what it does, where the result goes, what it should ask you first. A build goes the same way every time:

1. You describe it. Name the site, the API or the app if you know it.
2. Synapse AI researches what it doesn't know (below), and asks you, in one form, what only you can answer.
3. It shows The plan, in plain words.
4. You approve it, or say what to change, as often as you like.
5. It builds: it puts up a Connect card for each key the build needs, makes the agents and tools the plan needs, saves the workflow and checks it.

Not every workflow gets an agent. Where plain steps do the job (call a tool, ask a model once, wait for a person) the plan uses them, and says no agents: it runs on tools only, with no model called when no model is called at all. An agent goes where people ask it things as they come, or where the work needs judgement.

Research

Before it plans, Synapse AI reads each site or API your message names (up to three), once a conversation: the page's title, what it says it is, and any API spec it links to. A page that is only JavaScript is fetched with the scraper or a Browser step instead of an HTTP tool.

While it plans, it can stop to look things up, up to four questions a round:

• How SynapseOrch does something is answered from the product's own help and docs.
• A public fact (a site's pages, an API's sign-in or limits, who provides a service) is read on the web, within 6 web reads a question and 24 a plan.
• What only you know (your process, your customers, a value of yours) is asked of you, in one form.

The answers, with their sources, go into the plan.

The plan

The plan opens with You asked for, your request in its words, and then What it will do: one numbered row per step, coloured by who does the work (a model, logic, or a person). Under them it says when it runs, any cap on a run, the agents it makes, and what a run costs at most. When it applies, the card adds:

• New agents' instructions, for each agent it will make: these are covered by your yes.
• Asks you first and Acts without asking: what will wait for a person, and what will not. A tool that asks each time offers Let … act without asking.
• Models: the model each new agent gets and why, with a stronger one at its price as a button.
• After your yes: the Connect cards it will put up once you approve, one per provider.
• Missing or off: a tool or app the workspace does not have, or a feature that is off. It offers Turn on …, Build without …, or Find who can, which looks for a provider.
• Changed from the last plan, after you asked for a change: what is different, so a swapped site or tool is seen before your yes.
• Open questions and Not carried over.

Press Build it, Change something (it redrafts from what you type), or Stop instead.

After you approve

Synapse AI makes the agents and tools the plan needs, saves the workflow, checks it, and says what it did. A schedule is made paused: check what each run starts with, then turn it on under Schedules, or ask Synapse AI to. A workflow that never passes its check is removed again, with its first three issues: Removed the workflow "Weekly digest" again: it never passed validation: … An agent the build made for it, which nothing else uses, goes with it.

While Knowledge is on, Synapse AI knows the workspace's knowledge bases and can add Retrieve and Ingest steps.

Example requests

Build me a pipeline that takes a URL, scrapes the page, summarises it in three bullet points,
and saves the summary to the Vault as summaries/<the page's title>.md

Every weekday at 8 AM UTC, summarise yesterday's support tickets and post it to Slack

Search our shop for a product by name and add the first match to the cart

Change the research workflow to ask me before it sends anything

Asking for something now

Synapse AI explains the product and builds; it does not look things up for you in the chat. Ask what's the weather in Chennai and it offers a workflow that fetches it, once now or each morning, and asks where to send it. Say yes and the build starts from there. Ask what can you do for me and it answers with what the product can do, with an example of each.

HTTP tools from a URL, a curl, a spec or a docs link

Give Synapse AI what you have about an API, and the build makes HTTP tools from it:

• A spec (an OpenAPI file's URL, or its path in the Vault): it reads the spec as the Tools screen's import does, and each operation the plan uses becomes a tool, with the spec's address, inputs and sign-in. A spec read once is the conversation's: a later change in the same conversation can use its other operations.
• A docs link or the API's address: it reads the page and follows a spec the page links to. Given only the API's base address, it tries the usual spec addresses there. With no spec found, it asks you for the spec or the docs.
• A curl you paste: the tool takes the curl's method, address, headers, body and inputs as they are; only what varies from call to call becomes an input. The key in it stays out of the tool: see Keys.
• A URL and what it does: a single call, made as a new tool and checked as the Tools screen checks one.

A change to an HTTP tool shows its address before and after (an HTTP tool: GET … → GET …) and the inputs a new address adds. It never changes the tool's headers or its key.

Fixing a Python tool

Ask Synapse AI to fix a Python tool ("it crashes on an empty list") and the change shows the fix as a diff: removed lines in red, added lines in green. Before you see it, the fix is run on one to three sample inputs, with no internet, as the Tools screen's Test runs it:

• It passes: the review shows the results, and Go ahead saves exactly the code that passed.
• It fails: Synapse AI tries once more; if the fix still fails, it says The fix I wrote for "…" fails its test with …: … Nothing was changed.
• It needs the internet: a card, Test this fix to "…", runs it with the internet when you press Test this fix. Nothing changes until you press Go ahead after it passes.

Undo puts the tool's code back as it was, unless that code held a key, which was not kept.

Keys and sign-ins

Keys go on Connect cards, never in the chat. Synapse AI never asks you to type a key, a token or a password into the conversation. When a build needs one, it says where it goes and puts up a Connect card: an app's own sign-in, a token field, a model provider's key, an HTTP tool's header, a remote MCP server, or a link to the screen where it is kept. You type the key into the card yourself and press Connect; it goes to the product's own screens, and Synapse AI hears only Connected ….

A key pasted into the chat anyway is taken out before anything reads it, shown as ‹credential removed›, and Synapse AI says: I took a key out of your message before reading it … Please don't paste keys in this chat again: if it was real, rotate it at the provider, then connect the service on its Connect card, or add the new key in Settings → Vault, where it is kept encrypted.

A spec whose sign-in a card can't keep (OAuth or OpenID Connect) is imported on the Tools screen instead, and its key is set there.

See how I'm built

Synapse AI is an ordinary SynapseOrch workflow, with ordinary sub-agents. See how I'm built, in the panel's header, opens How I'm built: its workflow, large, on a read-only canvas, and under it where it starts, where it stops for you, its sub-agents and the step types it uses. Self-hosted, Duplicate to my workspace copies it into your workflows and opens the copy, to read or change.

Who can do what (Cloud)

Every member can ask Synapse AI and read its answers.

• A viewer asks and reads.
• An approver can also answer what waits on them and allow or refuse a tool call.
• A builder, admin or owner can also run and stop things, change and delete, and build.

A request beyond your role is answered before anything is promised: Not done: building and changing workflows, agents, tools, skills and schedules needs the builder role … Someone with the builder role can make this change. A card is pressed by a person, in the app. In the audit log, a change Synapse AI made carries what confirmed it: token for a card you pressed, or review: and the change's digest for a change you approved.

Self-hosted

• A new install's default agent is Synapse AI. An install whose old default agent was never edited has it turned into Synapse AI at its next start, under the same id, so its conversations stay; an edited one stays yours.
• Synapse AI is not listed on the Agents page and can't be put on a widget.
• Settings → Models → Builder model is its model, for every step of a build except the quick steps, which Model for quick steps sets. Reasoning-capable models do noticeably better.

Which model to use

Synapse AI does its best work on a capable model. Use Claude Haiku 4.5 or a stronger model: SynapseOrch 2.0's release tests built and ran every test brief with Synapse AI on Claude Haiku 4.5. Reasoning-capable models do noticeably better; a smaller model may plan less well, ask more, or miss a step.

• The picker under the message box, Model for this build (Builder model on Cloud), sets the model for this conversation; self-hosted, Set as default keeps it as the builder model.
• Settings → Models → Model for quick steps is the model for deciding what you asked for and whether you approved: a small, fast model usually does those as well.

Images and stopping

• Images: attach screenshots or diagrams to a message when the model can read images.
• Stop: while Synapse AI works, the send button becomes Stop Synapse AI. A build keeps running on the server if you close the panel or the tab, and a reload brings back a conversation that is still open.

Over MCP, the same reads and acts are tools of their own, and a delete there still takes its confirm_token.