Probe0 vs Cloudflare AI Gateway
Cloudflare AI Gateway is a hosted edge endpoint your applications call instead of the provider. Probe0 is a proxy that runs on your own machine and sits under the coding agents you already have open, so the traffic it inspects never leaves the laptop unless a cloud model is actually needed.
A hosted control plane on Cloudflare's edge that sits in front of 15+ model providers and adds caching, retries, fallback, rate limits, spend limits and per-request logs.
| Feature | Probe0 | Cloudflare AI Gateway | Edge |
|---|---|---|---|
| Where your prompts go | Stay on the Mac unless a cloud model is needed | Every request routes through Cloudflare's edge | Probe0 |
| Setup for one developer | One proxy and one cert, then Claude Code, Codex and Cursor are covered | Change the base URL inside each calling app | Probe0 |
| Local model routing | Ollama and LM Studio, weak answers retried on cloud | Workers AI runs on Cloudflare hardware, not yours | Probe0 |
| What a cache hit costs | Disk lookup, no socket opened | Edge cache shared across every caller and region | Even |
| Semantic cache | Local vector index with a strict floor, refuses tool-call traffic | Cache lives at the edge, no index on your machine | Probe0 |
| When the cap is hit | Warns, then pauses the run | Returns 429 or reroutes | Probe0 |
| Per-process attribution | Ledger names the process behind each call | Metadata tags you set in your own code | Probe0 |
| Knowing which feature saved the money | Each module toggles alone and reports its own savings | Request logs and cost analytics, no per-feature figure | Probe0 |
| One endpoint for many machines | One Mac, stops when the laptop sleeps | Managed global endpoint every machine and service can share | Competitor |
Cloudflare owns the case where many machines and people have to be governed from one place, and nothing on a laptop competes with that. My problem was narrower. I have three coding agent CLIs on one Mac, they burn tokens all day, and none of them are an application whose source I control. Repointing an SDK base URL is a fine adoption story when you wrote the code doing the calling, and a much worse one when the caller is Claude Code, Codex or Cursor, each with its own config surface and its own opinions about endpoints. Probe0 installs one local proxy and one certificate, and after that every agent on the machine is covered without touching any of their settings. Because the proxy is local, an exact cache hit is a disk lookup that never opens a socket, and the semantic cache is a vector index on the same disk that refuses to match any conversation carrying tool calls. I would rather lose the hit rate than let a near-miss corrupt a run that is editing files.
The rest of the modules follow from being on the machine. Local Routing can hand a request to a model already loaded in Ollama or LM Studio, and if the answer comes back weak it retries on the cloud, which is only possible when the proxy and the GPU are the same computer. Model Tiering tries the cheap model first and escalates when it has to. Request Coalescing collapses simultaneous identical calls into one upstream. Spend Guard caps a run or a day, warns, then pauses instead of handing an agent mid-run a 429 it will try to interpret. Every module is individually switchable and each reports what it saved, so you can turn one off and watch the number move. The Recording ledger stores model, tokens, real cost, latency and which process made the call, and because it knows your actual usage it can tell you when you are paying for a subscription tier above what you use.
Where Cloudflare is genuinely better, it is better for reasons a local proxy cannot argue with. Its spend limits are denominated in dollars, scoped by model, provider or custom metadata tag, and enforced centrally no matter which machine sent the request. Its cache is shared, so a hit paid for by one caller in Frankfurt serves the next caller in Sydney. Retries and cross-provider fallback happen in infrastructure that stays up when your machine does not, logs are durable and exportable through Logpush, and identity-aware budgets tied to Cloudflare Access are in closed beta for per-employee governance. Core features are free on every Cloudflare plan, with limits on log volume rather than a per-call fee. Against that, my honest limits: macOS only, private beta, sign-in through Google or GitHub with a young account system, and no team deployment at all. Everything Probe0 does, it does for one developer on one machine, which is exactly the bill I was trying to read.
The verdict
If the calls come from applications you deploy, or more than one machine or person has to be governed from a single place, use Cloudflare AI Gateway. For one developer on a Mac whose spend comes from coding agents they did not write, Probe0 is the better tool: one certificate covers every agent CLI, local routing and local caching keep work off the network entirely, the cap pauses a runaway run instead of alerting about it, and the ledger tells you which process spent the money. That is the setup I have, and Probe0 is what I run on it.
Questions
- Is Probe0 a self-hosted Cloudflare AI Gateway alternative?
- Sort of, but the shape is different. Probe0 is a local LLM proxy for one machine rather than a service you deploy for a team, so it covers a single developer's AI coding agent traffic and has no multi-user or multi-machine mode.
- Can Cloudflare AI Gateway route requests to a model on my own laptop?
- No. Its nearest equivalent is Workers AI, which runs models on Cloudflare's own edge hardware. If you want local LLM routing to Ollama or LM Studio with an automatic cloud retry when the local answer is weak, that has to happen on the machine the model is running on.
- Do I have to reconfigure Claude Code, Codex and Cursor separately?
- Not with Probe0. You install one local proxy and one certificate, and every coding agent CLI on the Mac routes through it. Cloudflare AI Gateway works by changing the base URL of the calling code, which is straightforward for your own apps and awkward for third-party agent CLIs.
- Which one gives better LLM cost control?
- For an organisation, Cloudflare: its spend limits are dollar-based, scoped by model or provider or tag, and enforced centrally. For one developer, Probe0 caps spend per run or per day and also records real cost per call with the originating process attached, which is the level of detail you need to work out where a single agent session went wrong.