Every capability, every safeguard
It reads your product. Then it runs it for your users.
No spec to write, no catalog to upload, no integration project to schedule. One line of setup, then it works out what your product can do and asks you which parts it may touch.
A support chatbot
Reads your help center and tells the person where the button is. Somebody had to write the help center, and somebody has to keep it true.
Syncanix
Reads the product itself and presses the button, in the person’s own account, with their own permissions, after showing them exactly what will change.
The four that matter
Four things you have to see to believe
Each one is hard to believe from a sentence, so each one is shown happening, exactly the way it runs in your product.
Nobody has to explain your product to it
It looks at the screens your product is made of, the forms, the buttons, the links, the things a person can fill in, and writes down what each one appears to do. Your engineers do not stop what they are doing, and your product does not change by a line.
- Mapping covers the first pass in one sitting: you click through your own product, signed in as yourself, and it writes down each screen as you go.
- It keeps reading afterwards as your own users move around, so a screen you ship on Tuesday is known about on Tuesday. Nobody files a ticket to teach it.
- If your engineers want it exact, a one-off scan of your source code fills in what a screen cannot show. It is optional, and nothing else waits on it.
It shows you the change before it makes it
This is the part that makes it safe to let an agent near a real account. Anything that changes data stops and shows a card with the actual before and after, the values themselves rather than a summary of them. Nothing runs until somebody presses the button.
- An approval covers one action, once. It cannot be reused for a second attempt, and it expires on its own.
- The agent acts as the signed-in person and inherits their permissions exactly. There is no service account behind it with a wider view.
- Every action lands in an audit trail you can read and export, and the trail is tamper-evident: a row altered after the fact shows up as one.
Nothing is switched on until you switch it on
Everything it finds arrives switched off, described in plain language, “upgrade a subscription” rather than an address on your API, so the person approving does not have to be the person who built it. You turn on what you want it to handle, and leave the rest.
- Scope it per capability, per group of users, and per page, so one part of your product can have a narrower set than the rest.
- Anything it discovered by looking at a screen always asks the person before it runs, whatever else the workspace has configured. That one is not a setting.
- Switch off one capability, one surface, or the whole agent from the dashboard. It takes effect on the next request. No release, no ticket.
It shows up where your users already are
The same catalog serves two places: a chat surface inside your own product, and a connection your customers can add to Claude, Cursor or ChatGPT and drive your product from there. One set of approvals, one audit trail, no second thing to configure.
- The chat surface reads your fonts, colors and corners off the page it lands on, so it looks like your product from the first load.
- Each workspace gets its own connection and its own credential, so one workspace can never reach another’s.
- When the right answer is a screen the person never found, it takes them there and highlights the control rather than pressing it. Having it done for them is the other path, and it runs through the same approval card.
Proof
Numbers, not adjectives
Each of these is measured by a test that fails the build when it stops being true.
- 6
- languages, right-to-left included
- < 30 KB
- added to a page load
- EU only
- storage and AI, in Frankfurt
- 12
- logins it plugs into
Fourteen capabilities · four jobs
What ships when you embed Syncanix
Grouped by the job each one does for you, and every claim carries its evidence: a number, a standard, or somewhere you can go and check.
It reads your product
Nobody writes a specification. It maps what you have already shipped, then reads it again every time you ship something new.
Finds what your product can do
It works out what your product can do by looking at it, at the screens, the forms and the buttons, while your engineers carry on with something else. If you want it exact, an optional one-off scan of your source code covers 39 frameworks across 19 languages at an extraction accuracy of 85% or better.
Catches changes before your users do
When your team moves a screen or renames a field, the agent notices and flags it, before a customer asks it to do something that no longer works. An optional add-on called api-witness checks what it believes against what your product actually serves.
In your app, and in their AI tools
The same capability catalog drives an embeddable widget and a standards-compliant MCP server. Your customers chat from your app; their tools (Claude Desktop, Cursor, ChatGPT) connect over MCP. One source of truth, two distribution surfaces.
No single AI vendor to depend on
Anthropic, OpenAI, and AWS Bedrock-EU wired in on day one. Failover routing per request, cost-aware model selection, and no single-vendor lock-in.
It acts as your user
Every action runs with the permissions of the person asking, is shown to them before it commits, and is written down afterwards.
Acts as the signed-in user
The agent acts as the signed-in user, never as a shared bot with the run of the place. The model never holds your users’ credentials, and every action is authorized and audited under that user’s own identity and permissions.
Works with the login you already have
Your users keep the login they already have, so you do not migrate anybody. It plugs into Auth0, Clerk, Cognito, WorkOS, Firebase, Supabase, Keycloak and Microsoft Entra directly, and into anything else through the open standards: 12 adapters out of the box.
Risky actions need a confirmation
Before anything changes, the person sees the exact change and has to approve it. That approval covers one action, once. It cannot be replayed, and it expires on its own. Sensitive actions can additionally require a fresh sign-in, using the RFC 9470 standard.
Audit trail by default
Who asked, what ran, what changed, and when, for every action, every approval, and every admin change. Nothing has to be switched on. The record cannot be quietly edited, and it is filed against the SOC 2 criteria an auditor asks for: CC6 (access) and CC8 (change). Rows are kept 13 months, and security events are copied to a 7-year write-once archive.
It looks and speaks like you
Your type, your colors, your six languages, on every device, and light enough that nobody notices it arriving.
Widget Design Studio
Author your agent’s look in the Widget Design Studio: color, type, and shape mapped to your product. It picks up a matching theme from shadcn, MUI, Chakra, and Mantine, so the widget is themed to your design system in minutes, not sprints.
Six languages, right-to-left included
EN · ES · FR · DE · AR · HE at launch, and not only the interface: the agent’s answers, and the instructions behind them, are localized too. Arabic and Hebrew get true right-to-left layout, mirrored components, and native-speaker review every release.
Under 30 KB, and usable by everyone
It costs your page under 30 KB to load, and the chat itself only downloads when somebody opens it. It cannot clash with your styles, it works from the keyboard, screen readers announce it, and it reads correctly right-to-left on any device.
It proves what it did
Where your data lives, what changed between one version and the next, and a way to check every action from inside your own systems.
Hosted in the EU
All customer data is stored in the EU (Frankfurt), and default AI inference runs there too, on Claude via Amazon Bedrock in Frankfurt, never used to train a model. Retrieval stays in-region. Encryption at rest with a separate key for every customer, held by you on Enterprise; EU-resident authentication; and you can bring your own AI provider keys on every tier.
Quality tests before anything goes live
Before a change to the agent reaches your customers, it is tested against your own past conversations. Is it still accurate, does it still pick the right action, does it still refuse what it should. A drop of more than 5 percentage points blocks the release, and an override has to be written down.
Your own systems can verify every action
If your engineers want a second lock on the door, your own systems can independently check that an action really came from an approved request, in 8 languages, each in its own native toolchain, so nobody has to learn ours.
Fourteen capabilities, one line of setup.
We point it at a staging copy of your product and show you the list it comes back with.