01
LLM gateway
Multi-provider routing. Agent traffic is matched to a model on cost and capability.
Product
Gatehold
In active development
AI agents draft infrastructure and code changes inside your own cloud account, and gates decide what ships.
Gatehold is agentic infrastructure in active development. Agents draft infrastructure and code changes inside the cloud account you already run. Policy gates and human approval decide what ships. It currently runs on our own infrastructure.
Components
Three components.
An LLM gateway, MCP tool infrastructure, and rails and gates.
02
MCP tool infrastructure
Ephemeral, self-cleaning compute that MCP tool servers run on, with automated reaping of idle resources.
03
Rails and gates
Workload identity for each agent, policy gates that fail closed, and human approval.
Flow
How it works.
The gate and a person sit between a draft and a ship.
-
01
Agent drafts
An agent proposes an infrastructure or code change inside the account.
-
02
Policy gate checks
The gate evaluates the draft and fails closed.
-
03
Human approves
A person approves the change.
-
04
Change ships
The approved change ships.
Architecture
Built on.
The architecture being built uses these Cloudflare pieces. Gatehold is in active development.
- Cloudflare Workers Gateway runtime for the LLM proxy.
- Durable Objects Per-agent session state and rate limits.
- R2 Run logs, artifacts, and audit trails.
- Workers AI Inference on open models alongside external providers.
- Queues Gated, human-approved change pipelines.
Writing
From the writing.
Longer write-ups of the same ideas.