Replicate MCP Server
Solid all-round agent support with no single blocker.
The widest model catalogue reachable over MCP, priced per compute-second. Cap spend before handing the key to an autonomous loop.
Agent-readiness
Good overall
Weighted across four pillars
What an agent can do here
Official MCP reference in Replicate's docs — agents run thousands of community and first-party models by prediction.
- Action depth
- Core write
- Hosting
- Self-run
- Auth
- API key
- Unattended
- Headless-capable
- Blast radius
- Medium
Agents can do the product’s core job — edit the timeline, ship the deploy, update the record.
You run the server process yourself.
A long-lived token works, so this can run on a schedule.
Mistakes are recoverable but visible.
What it costs to let an agent act
FreeUsage-meteredNo plan gate — billed per second of model compute. An agent looping on video models can spend real money fast.
How the score breaks down
Every factor below is scored by a fixed rubric, not by opinion — so two tools with the same shape always get the same number.
Reach
Can an agent actually do the job?
- MCP serverFirst-party2/2
- Action depthCore write1/1.5
- HostingSelf-run0.5/1
- Fallback pathPublic API0.5/0.5
Cost
What does it cost to let the agent act?
- Plan requiredFree3/3
- Usage meteringUsage-metered1/2
Control
Can I let it run unattended?
- Permission scopingAll or nothing0/1.5
- Unattended authHeadless-capable1.5/1.5
- Blast radiusMedium1/1.5
- Auth frictionAPI key0.5/0.5
Trust
Will this still work next month?
- MaintainerVendor2/2
- MaturityGenerally available1.5/1.5
- Research confidenceMedium0.5/1
- FreshnessVerified 2d ago0.5/0.5
Replicate vs other ai platforms tools
Sorted by overall score. Filter and sort the full category →
Reviews
Write a review
Please log in to leave a review
Log InSetup
How we know
- Last verified
- 2026-09-29 · 2d ago
- Research confidence
- medium
- Maturity
- Generally available
Spotted something wrong? Send a correction — vendor corrections are applied first.