Connectors · AI
Replicate MCP server for AI agents.
actionlake is an MCP gateway: one remote MCP server that connects Claude and other MCP clients to Replicate and the rest of your systems. It gives your agent 33 Replicate actions (19 reads, 8 writes and 6 that wait for a confirmation) and holds the Replicate credential itself, so the agent never does. It paces Replicate's limits and records every call.
100,000 free calls a month. No card, no email, no password: press Start a free sandbox on the next screen and keep the key it shows you.
claude mcp add --transport http actionlake https://connect.actionlake.com/mcpdo { action: "replicate.collections.list", intent: "check which Replicate account is connected"}# List collections of models · read · 1 upstream call- 33actionsin one signed file
- 19readlook, never change
- 8writecreate and update
- 6wait for a yesdestructive, confirmed first
Set up
Connect Replicate in three steps.
- 01
Add actionlake to Claude.
Run the command once in Claude Code; its consent page can create a free sandbox: 100,000 calls a month, no card. In the Claude app, add a custom connector with the URL below and paste your key on the consent page. Creating a new sandbox from the Claude app is not yet tested by us.
https://connect.actionlake.com/mcp - 02
Connect Replicate.
Bearer token. No consent screen: your agent can connect it for you, but the key then passes through your chat. To keep it out of the model's context, paste it in the console instead. actionlake stores what you paste write-only, sealed under your tenant's key.
api_tokenReplicate API token (starts with r8_), from replicate.com/account/api-tokens
account { op: "connections.requirements", service: "replicate" } # free: exactly what to paste, and what is found for you - 03
Make the first call. Check before a write.
Ask in plain words; Claude picks the action. Before anything changes,
checkvalidates the call and prices it, for free.do { action: "replicate.collections.list" }# list collections of models · read · 1 upstream callcheck { action: "replicate.models.create", input: { … } }# free: inputs, permissions, cost and confirmations, nothing sent
Replicate at a glance
- Sign-in
- Bearer token. No consent screen: your agent can connect it for you, but the key then passes through your chat. To keep it out of the model's context, paste it in the console instead.
- You provide
api_token- Limits we pace to
- 3,000 calls a minute, per connection
- API covered
- Replicate HTTP API v1 (OpenAPI 1.0.0-a1)
api.replicate.com - Built from
- Replicate's documentation, read 23 October 2026api.replicate.comreplicate.com
- Service file
- Version 1.0.0, signed (Ed25519)
- Cached reads
- 18 of 19 reads, for up to 5 minutes; a write clears what it affects
- Shared entities
- 1, so ask can line Replicate up with your other systems
- Data
- Processed and stored in Sydney, Australia. No US region yet
Capabilities
What an AI agent can do in Replicate.
33 actions in 14 areas. Every one carries its effect, so a key can allow the reads and hold the rest.
- 19readlook, never change
- 8writecreate and update
- 6destructivewait for a confirmation
Deployments
52 read · 2 write · 1 destructive
List deployments · Get a deployment · Create a deployment · Update a deployment
Models
52 read · 2 write · 1 destructive
Get a model · Create a model · Delete a model · Update metadata for a model
Predictions
42 read · 1 write · 1 destructive
Get a prediction · List predictions · Create a prediction · Cancel a prediction
Trainings
42 read · 1 write · 1 destructive
List trainings · Get a training · Create a training · Cancel a training
Files
32 read · 0 write · 1 destructive
List files · Get a file · Delete a file
Model versions
32 read · 0 write · 1 destructive
List model versions · Get a model version · Delete a model version
Collections
22 read · 0 write · 0 destructive
List collections of models · Get a collection of models
Account
11 read · 0 write · 0 destructive
Get the authenticated account
6 more areas · 6 actions
- Deployment predictions 1
- Hardware 1
- Model examples 1
- Model predictions 1
- Search 1
- Webhooks 1
Top actions
The Replicate API, shaped for agents.
The ids your agent types, what each does, and the one name it shares with other systems. 8 of Replicate's 8 action names are shared with at least one other connector, so an agent learns them once.
| Action | What it does | Effect | Same name in |
|---|---|---|---|
replicate.account.getGet the authenticated account | Get the authenticated account. | read | identity.current_user.get307 connectors |
replicate.models.getGet a model | Get a model. | read | infra.ml_model.get13 connectors |
replicate.search.querySearch models, collections, and docs (beta) | Search models, collections, and docs (beta). | read | infra.ml_model.search3 connectors |
replicate.predictions.getGet a prediction | Get a prediction. | read | infra.ml_task.get3 connectors |
replicate.models.createCreate a model | Create a model. | write | infra.ml_model.create5 connectors |
replicate.models.deleteDelete a model | Delete a model. | destructive | infra.ml_model.delete9 connectors |
replicate.models.updateUpdate metadata for a model | Update metadata for a model. | write | infra.ml_model.update5 connectors |
replicate.model_predictions.createCreate a prediction using an official model | Create a prediction using an official model. | write | infra.ml_model.run3 connectors |
replicate.files.listList files | List files. | read | this connector |
replicate.models.listList public models | List public models. | read | this connector |
replicate.hardware.listList available hardware for models | List available hardware for models. | read | this connector |
replicate.trainings.listList trainings | List trainings. | read | this connector |
Try it
Prompts to paste into Claude.
Each one names the Replicate actions it will call, so you know what it reads and what it changes before you send it.
- Readread
List collections of models in Replicate and tell me what stands out.
calls
replicate.collections.list - Writewrite
Create a model in Replicate. Show me exactly what you will send before you save it.
calls
replicate.models.create
Governance for Replicate
Destructive actions wait for a yes.
6 Replicate actions are destructive. These are all of them:
- Delete a model
replicate.models.delete - Delete a file
replicate.files.delete - Delete a deployment
replicate.deployments.delete - Delete a model version
replicate.model_versions.delete - Cancel a training
replicate.trainings.cancel - Cancel a prediction
replicate.predictions.cancel
A destructive action runs only with confirm: true. Under the human policy the agent cannot confirm for itself: the run stops until a named person approves, in the console or from a signed link in an email. The cap on a key is refused before the call is sent, never after.
Delete a model?
intent: the user asked me to delete a model
account { op: "keys.create", name: "replicate-agent", allowed_actions: ["replicate.*"], cap_calls: 2000, confirm_policy: "human_for_writes", approval_rules: [ { match: "replicate.models.delete", approver: "[email protected]" }, { match: "replicate.files.delete", approver: "[email protected]" }, { match: "replicate.deployments.delete", approver: "[email protected]" } ]}Cost
What Replicate costs on actionlake.
Every Replicate action is one upstream call, the only thing actionlake bills. Searching, check, cost, account and any call Replicate refused with 401, 403, 429 or a 5xx are free. Paid plans start at US$49 a month for 250,000 calls, with overage from US$0.30 down to US$0.18 per 1,000.
- 1upstream call per actionall 33 Replicate actions
- 8actions pageeach page of a long list is one call
- 18reads can be cacheda cache hit costs nothing
- 100,000calls a month, freepaid plans from US$49 a month for 250,000
Action index
Every Replicate action.
33 actions by area: the id your agent calls, what it does, and its effect.
Deployments5
replicate.deployments.createCreate a deploymentwritereplicate.deployments.deleteDelete a deploymentdestructivereplicate.deployments.getGet a deploymentreadreplicate.deployments.listList deploymentsreadreplicate.deployments.updateUpdate a deploymentwrite
Models5
replicate.models.createCreate a modelwritereplicate.models.deleteDelete a modeldestructivereplicate.models.getGet a modelreadreplicate.models.listList public modelsreadreplicate.models.updateUpdate metadata for a modelwrite
Predictions4
replicate.predictions.cancelCancel a predictiondestructivereplicate.predictions.createCreate a predictionwritereplicate.predictions.getGet a predictionreadreplicate.predictions.listList predictionsread
Trainings4
replicate.trainings.cancelCancel a trainingdestructivereplicate.trainings.createCreate a trainingwritereplicate.trainings.getGet a trainingreadreplicate.trainings.listList trainingsread
Files3
replicate.files.deleteDelete a filedestructivereplicate.files.getGet a filereadreplicate.files.listList filesread
Model versions3
replicate.model_versions.deleteDelete a model versiondestructivereplicate.model_versions.getGet a model versionreadreplicate.model_versions.listList model versionsread
Collections2
replicate.collections.getGet a collection of modelsreadreplicate.collections.listList collections of modelsread
Account1
replicate.account.getGet the authenticated accountread
Deployment predictions1
replicate.deployment_predictions.createCreate a prediction using a deploymentwrite
Hardware1
replicate.hardware.listList available hardware for modelsread
Model examples1
replicate.model_examples.listList examples for a modelread
Model predictions1
replicate.model_predictions.createCreate a prediction using an official modelwrite
Search1
replicate.search.querySearch models, collections, and docs (beta)read
Webhooks1
replicate.webhooks.get_default_secretGet the signing secret for the default webhookread
Questions
The Replicate MCP server, answered.
Is there a Replicate MCP server for Claude?
Yes. actionlake is an MCP gateway: a remote MCP server that gives Claude, or another MCP client, 33 Replicate actions across deployments, models, predictions and trainings, through the same eight tools it uses for every other connector. actionlake is independent; it is not made or endorsed by Replicate.
What do I need to connect Replicate?
Replicate API token (starts with r8_), from replicate.com/account/api-tokens. The free call account { op: "connections.requirements", service: "replicate" } lists exactly what to paste.
Can an AI agent delete or change things in Replicate?
Only if its key allows it, and only with an explicit confirmation. 6 Replicate actions are destructive, including delete a model, delete a file and delete a deployment. Each runs only with confirm: true. Under the human confirm policy, or an approval rule on the key, the run waits until a named person approves it.
Will an AI agent use up my Replicate API limit?
actionlake paces every call to the limits in the Replicate service file: 3,000 calls a minute per connection. Repeated reads are answered from cache (18 Replicate actions can be), which never reaches Replicate.
What does a Replicate action cost?
One upstream call, the only unit actionlake bills. A list that pages bills one call per page (8 Replicate actions page). Cache hits, searches, dry runs and calls Replicate refused with 401, 403, 429 or a 5xx are free. The Free plan includes 100,000 calls a month with no card. Paid plans start at US$49 a month for 250,000 calls, with overage from US$0.30 down to US$0.18 per 1,000.
Which Replicate API does it cover?
The Replicate HTTP API v1 (OpenAPI 1.0.0-a1) at api.replicate.com. It is built from Replicate's documentation as read on 23 October 2026. 33 actions in 14 areas, each one in the signed service file (version 1.0.0).
Does actionlake keep my Replicate data, and where?
actionlake processes and stores data in Sydney, Australia; there is no US region yet. Reads may be cached for up to 5 minutes, and a write clears the reads it affects. Each run's request and response is kept, sealed, for your plan's trace window: 1 day on Free, 3 days on Starter, 7 days on Pro, 14 days on Team, 30 days on Business and 30 days on Enterprise. Your Replicate credential is sealed under your tenant's key, never logged and never returned. Your traffic is not used to train models.
Which AI clients can use it?
Claude Code, tested by us end to end, and the Claude app as a custom connector, tested by us with an existing key. Other MCP clients that support remote servers over Streamable HTTP (MCP revisions 2025-06-18 and 2025-11-25) and OAuth dynamic client registration, or a bearer header, should connect the same way; we have not tested them. Or plain REST, under the same action names.
Give your agent Replicate today.
Start free - 100,000 calls/month, no card. No signup. Hard caps on every key. Cached answers are free.