Integrations/Replicate

Control what your AI agents do in Replicate.

Let your team build AI agents for Replicate. Bollard AI controls what each one can do, and anything that spends stays off by default.

Request early accessSee how it works →
Replicate
Inference Agent
Models Account areaOffReadFull use
Predictions Account areaOffReadFull use
Deployments Account areaOffReadFull use
Run predictions
What agents can do

Set what each agent can do across your account.

Each area of Replicate, like models, predictions, or deployments, gets one of three levels: Off, Read, or Full use. Anything not allowed is denied, and anything that spends is a separate switch.

Off
No access
The agent can't see that part of your account.
Read
Read-only
The agent can read models and past predictions in that area. It can't change or run anything.
Full use
Create and update
The agent can create and update model details in that area.

Some actions stay off until explicitly turned on

Each one spends money or removes work for good, so it could override the limits set elsewhere.

Running predictionsCreating deploymentsDeleting models
Set up & manage

Connect once. Manage everyone from there.

Connect once, then control access person by person. No new keys to hand out.

Step 1 · Set up

Connect once, company-wide

An admin connects your company's Replicate a single time. Bollard AI lays out the areas of your account, and anything you don't allow stays out of reach for every agent.

Step 2 · Manage

Control access per person and agent

From one dashboard, set what every person and their agents can reach. Change it any time, with no new keys for anyone.

OffNo access
ReadRead-only
Full useRead, create, and update model details in a single area
Example

A model catalogue agent.

An inference agent updates model details and can read past predictions, but it can't run a prediction or create a deployment.

Models Account areaOffReadFull use
Predictions Account areaOffReadFull use
Deployments Account areaOffReadFull use

Manage access

ALLOWEDInference Agent updated a record in Models
ALLOWEDInference Agent read a Prediction
BLOCKEDInference Agent tried to open Deployments, denied by default

Monitor activity

The security model

Top-tier security.

Bollard AI is access control for AI agents. It sits between your agents and Replicate, so they never hold your Replicate keys and never reach Replicate directly.

Everything runs through Bollard AI

Agents send their requests to Bollard AI, not to Replicate. Bollard AI checks each one against your rules and passes only the allowed requests through.

Denied by default

Every area starts off-limits, and anything that spends is held back on top of that. An agent gets in only where you've allowed it, and anything Bollard AI doesn't recognise is refused and held for review, never waved through on a guess.

Your keys stay sealed

Your Replicate keys are sealed in a vault unique to your company and never shown again. Bollard AI unlocks them only for the instant an allowed request needs it.

Activity you can read

Every attempt is recorded in plain English: who or which agent, what they tried, which model or prediction, and the decision. The record is readable and can't be edited after the fact.

Questions

Replicate, answered.

Does connecting Replicate hand Bollard AI your whole account?+
Only the areas you allow. An admin connects your Replicate, and you decide which areas, like models or predictions, each agent can reach. Bollard AI doesn't browse or copy your account. It sits in front of it and checks every request the agents make, passing or blocking each one.
Can an agent run predictions and spend money?+
Not unless you turn it on. Running predictions and creating deployments are held off by default, so an agent can manage model details and read past runs without ever spending money, until you explicitly allow it.
Can an agent see more than the person who owns it?+
No. An agent reaches only what its owner reaches. An explicit Off always overrides an inherited permission.
Where do our Replicate keys live?+
Sealed in an encrypted vault, behind a key unique to your company. It's never shown again once connected, and Bollard AI unseals it only for the instant an allowed request needs it.
What happens when an agent tries something it's not allowed to do?+
Bollard AI blocks the request before it reaches Replicate, and the blocked attempt lands in the activity record with the reason.
How do we connect Replicate?+
An admin connects the company's Replicate once. Bollard AI seals the connection and lays out the areas of your account to set rules for. From then on, all management happens in the Bollard AI dashboard.
Early access

Put your team's agents to work in Replicate, without the risk.

We're onboarding our first development partners now. Tell us about your setup and we'll be in touch.

Request early accessAll integrations →