a9script is an integration platform for the API integration and workflow automation of a business. It connects your systems and runs the automations between them. The logic between your systems lives in one place, runs there and can be read there. Every call to another system is configured once and recorded as it runs, and the code that remains is the business logic in plain JavaScript, written by your team or by an AI agent.
A failed run turns red until a person acknowledges it. After a restart, a run continues from its last saved point, and a call already made is never made twice. An endpoint can be offered as an MCP server, so an AI assistant works with your configured calls as tools.
One endpoint of a9script can be read in four ways. As a webhook it acknowledges the sender at once and runs your scripts afterwards. As an API the caller waits for the answer of one script. As an MCP server an AI assistant lists your operations and calls them as tools. As an app it serves a folder of files at your address with sign-in.
A legacy REST API becomes callable by an assistant without a server written for it. A webhook can carry several scripts, each with its own filter, so you decide which script reacts to which message. Every activated endpoint describes itself.
MCP is the open standard AI apps use to work with tools and data. The tools come with generated schemas, and the credential stays on the platform.
Every system you connect has one place with its address and its credential. A script never sees a secret: the platform attaches the credential, refreshes an expired token once and sends the call again.
Senders such as GitHub, Atlassian Jira, Microsoft Graph, Shopify, Slack and Stripe come pre-configured. Sandbox and production each have their own connections and credentials, and a promotion carries none. You enter a secret once, and no screen returns it.
The platform keeps the credentials encrypted at rest and attaches them to every outgoing call.
A call to another system is configured once: address, credential, retries, paging and what counts as an answer. You name it in your own words, and the script calls it by that name. Test shows the answer the script gets and the request that was sent, with the credential named and never shown.
The script editor completes every configured call with its signature and answer on hover, and marks a name that resolves to nothing while you type. Your team tests a call before it is activated.
Calls to HTTP APIs, mail, CSV files, mappings and AI models are configured the same way. A model call holds prompt, model and the connection with the key in one place, and the answer of a model is input the script validates.
A run is written down step by step. A transient failure is repeated, and a status such as a rate limit parks the whole run for the time the other system asked, up to 24 hours, holding nothing open. After a restart, a deployment or a lost machine the run continues from its last saved point, and a call already made is never made twice.
In the sandbox you start a run with trace and read every call, what it sent and what came back. A stepped run parks after every external call, so you can check each step before the next.
Transient failures are repeated three times by default. A sandbox run with trace keeps every request and answer for 7 days.
Senders such as Atlassian Jira, GitHub, Linear, Microsoft Graph, Okta, Shopify, Slack, Stripe, Twilio and Zoom come pre-configured. Registration handshakes are answered without a run, and ignore rules drop your own writes coming back.
Transient failures are repeated, and a rate limit parks the run for the time the other system asked, up to 24 hours. A night job that meets a limit at two in the morning is finished by three, and paging takes one line.
Thousands of records a night are read page by page, mapped, checked for duplicates and written in bounded pieces. Records are matched across systems by source id, and a bad row is set aside with its key and reason while the job continues.
A failed run turns its script red until a person acknowledges it, and a connection that meets faults in a row opens its circuit and turns red as well. Notification rules mail you, throttled, and system triggers start a script of your own.
Every activation is a numbered version with history, diff, blame and restore. Production is marked in colour, and activating or deleting there asks first. A promotion from the sandbox shows a plan before anything changes.
Callers of your endpoints are users of your environment, invited by mail through your own mail account and given the roles your environment defines. The script sees who called.
Every recorded number that informs a decision is drawn on a page that names its question and its decision. The pages hold counts, durations, codes and names, never the content of a message.
An AI agent builds integrations from a terminal or over MCP and never holds a credential.
Tenants and environments are sealed off from each other, and secrets are entered once and never shown again.
