AI pipelines that do the work,
and stop where you decide.

How it works

Four kinds of step. Any work you can describe.

A pipeline is a row of stages. A card moves through them, collecting what each step produces, and stops wherever a person should look.

1. Intake

   Anything that arrives lands here.

   A file drop, the API, a webhook, or a trigger like a new Gemini meeting-notes doc. Intake works out what it is, unpacks an archive into one card per file, labels it, and sends it to the pipeline whose entry gate accepts it. If none does, or several do, it waits for a person.

2. AI step

   Skills do the reading and writing.

   A skill is a prompt with a typed output: pull out the terms, compare them with yours, write the summary. An answer that breaks its output is retried. Each kind of step runs on the model you choose for it.

3. Gate

   People decide at the gates.

   A gate stops the card and asks a question with set answers: approve or reject, share or keep. Answer it on the card, the board, the dashboard or the inbox, one at a time or several together.

4. Tool

   Tools reach out. Results remember.

   Tools read a Google Doc, send Gmail as an account you connected, or run one you wrote. A pipeline keeps its results, so the next run can look up where a subject stood the last time it came up.

Run

Everything waiting on you, in one place.

The dashboard shows what every pipeline is doing and what it needs from you. Answer a gate right there, or open the card it's on.

• InboxEvery card waiting on a person, oldest first. Answer one, or tick several and answer them together.
• BoardsOne column per stage, live, with each stage's p50 time, its cards per day, and any step that is failing.
• DeliverablesWhat every run produced, by subject, with each metric's trend. Tick subjects to run the pipeline on them again.
• Inputs and schedulesSaved inputs start a run now or on a schedule. Webhooks and triggers, like a new Gemini notes doc, start them from outside.

On the record

Every model call, recorded with the step that made it.

Ergates records each call to its model in its own database: skill steps, intake's labels, the authoring chat. Each one is tied to its card, stage and skill, scored by a judge, and checked against detection rules.

• Observe firstRules record their verdicts and block nothing until an admin opts a rule in.
• Halt when it mattersIn blocking mode, a rule you allowed halts the session when it fires, and it stays halted until an admin releases it.
• Quality per stepThe dashboard shows each step's rule firings and judge scores. Improve hands the authoring chat a prompt to draft a fix.
• Aged out on a scheduleClean sessions are kept for 30 days. Anything a rule flagged is kept for 180.

Yours, not shared

One instance per organization.

Ergates doesn't pool customers into one database. Each organization runs its own instance, with its own data and its own people.

• Your own instanceYour pipelines, skills, cards and model log live in a database that belongs to your instance alone.
• Your Google, your OAuth appErgates connects through an OAuth app you create as Internal to your own Workspace, with two scopes: Drive read and Gmail send. A step names the account it works as, and only the pipelines that name it can use it.
• Secrets stay where they areType an API key in and it's sealed and never shown again. Or point Ergates at Infisical: it keeps only where the secret lives and reads it each time, so a rotation applies right away.
• Your modelConnect Claude, OpenAI, Gemini, Gemini on Vertex AI, or any OpenAI-compatible endpoint, and pick the model for each kind of step. Claude, OpenAI and Vertex can sign in with your workload identity, so no key is stored.
• Backup and restoreEvery definition is backed up daily. A restore shows what it will change before it changes anything.
• Sign-inGoogle, a passkey, or a password set from a one-time link. Removing someone ends their sessions at once.