Video wird geladen...
Video konnte nicht geladen werden
OpenAI just dropped their new Agents API 🔥. You can now run Codex in Cloudflare Containers. OpenAI's Intelligence with Cloudflare's security and isolation. In ~4 minutes we go from creating OpenAI keys to a fully deployed agent that stores all outbound traffic in D1.
24,282 Aufrufe • vor 19 Stunden •via X (Twitter)
23 Kommentare

Check it out ↓

Hey @Jilles would you like to give a talk at #Sailsconf2026? It’s all online and pre-recorded

Sure! Sounds awesome.

Awesome please submit a CFP!

Seeing the Codex harness running in Cloudflare containers is a huge milestone. We actually went deeper on this here:

the D1 outbound log is what people will copy. the container is just the fence

Cloudflare is for total losers. Cloudflare is anti-small business, anti- entrepreneurial, and anti-internet

L take Any small business can thrive on the free plan. Name a company with a more generous free tier

Your company is built on L's

四分钟跑通部署,出站流量进 D1 这点挺实用的

Logging outbound in D1 is the right instinct. What actually makes the container a boundary is a closed egress list. If it can still hit MCP, DNS, or a sidecar that never touches D1, the log is a souvenir.

The interesting part of running it in someone else's sandbox is what the agent can still reach from inside. Isolation is easy to demo and the real question is which credentials cross the boundary with it.

does the four-minute setup also cover rotating OpenAI keys and inspecting failed outbound requests in D1?

I'd keep the session ID and tool-call ID alongside each outbound request in D1. That makes 'why did the agent contact this service?' traceable back to a specific step when a run goes sideways.

4 minutes is fast, but D1 writes on every outbound call is the part that'll bite at scale. D1 tops out around 10GB per database and read replication lags under load, so a chatty agent logging full request bodies fills that faster than people expect.

> How do you know which agents can access your internal systems right now?

The four-minute path matters. For GTM agents, logging every outbound call in D1 makes sequence bugs far easier to find than chasing model logs.

4 minutes to deploy is fast

Can we have them use subscriptions?

Tried this shape last month with a Claude agent on a VPS. Isolation kept breaking when the agent wrote outside the allowlist. Cloudflare containers plus a traffic log in D1 is the missing half I kept hacking with shell wrappers.

4 minutes to deployed is fast but Cloudflare D1 storing all outbound traffic is the part worth stress-testing before anything sensitive goes near it

#cloudflare 👀

@lucas_montano, a @OpenAI vai comprar a @Cloudflare ?








