Probe0 vs Vercel AI Gateway

Vercel AI Gateway is infrastructure your application calls over the network. Probe0 is a proxy that runs on your laptop and sits under the coding agent CLIs you already have open.

A hosted, Vercel-operated routing layer that gives applications one endpoint into a large catalog of model providers, with team budgets, failover and a usage dashboard.

FeatureProbe0Vercel AI GatewayEdge
Where your prompts goStay on your Mac. No Probe0 server exists.Through Vercel's network. Hosted only, no self-host.Probe0
Setup for one developerOne proxy plus one certificate. Claude Code, Codex and Cursor all route through it.Change your app code, usually through the AI SDKProbe0
Local model routingSends work to Ollama or LM Studio, retries weak answers on the cloudRoutes to hosted provider APIsProbe0
Cost of a cache hitZero. Exact hits are served from local disk, no network call.Prompt-similarity caching inside the hosted layerProbe0
Cache safety on tool callsRefuses semantic matches on anything carrying tool callsSimilarity caching, matching rules not published in detailProbe0
What happens at the capWarns, then pauses the run. Per run and per day.Budgets per key, project and team. Returns HTTP 402 when exceeded.Probe0
Who spent the moneyLedger records model, tokens, real cost, latency, and which process calledDashboard and Custom Reporting API by user, tag, model, providerProbe0
Subscription tier adviceReads real usage, says when your tier is above what you useReports token usage, does not advise on your planProbe0
Model catalog and provider failoverOnly what coding agents call. No cross-provider failover.Large multi-provider catalog across text, image, video and voice, with automatic failoverCompetitor

Vercel AI Gateway's catalog and automatic failover are genuinely strong, and if a provider rate-limits you mid-request it will quietly try another one. That matters for an application serving other people. It is not what was draining my account. My spend came from Claude Code, Codex and Cursor running on my own laptop, in loops I could not see into, and a hosted gateway cannot reach that traffic because the agent CLI picks its own model, its own retries and its own context and never routes through infrastructure I control. Probe0 installs as a local proxy with a certificate installed once, and from that point every coding agent CLI on the machine goes through it with no per-tool configuration.

Being local buys capabilities a hosted service structurally cannot offer. Probe0 sends a request to a model already loaded in Ollama or LM Studio and only falls back to the cloud when the answer comes back weak. Model Tiering tries the cheap model first and reaches for the expensive one only when needed. The exact cache lives on local disk, so a hit costs nothing and makes no network call at all. The semantic cache is a local vector index behind a strict similarity floor, and it refuses to match anything carrying tool calls, because one wrong hit in the middle of an agent run corrupts the run and costs more than it saved. Request Coalescing collapses simultaneous identical calls into one upstream request.

Every module switches on and off independently and reports what it saved on its own line, which is how I would rather you judge the numbers than by taking my word. Spend Guard caps per run and per day, warns first, then pauses, so an overnight loop stops instead of billing until morning. The ledger records model, tokens, real cost, latency and the process that made each call, and that per-process record is also how the plan-tier advice works: it sees your actual usage and tells you when you are paying for a subscription tier above what you use. Sign-in is Google or GitHub, no password. The honest limits: macOS, private beta, one developer per machine, and provider coverage scoped to what coding agents call rather than a hundred-provider catalog.

The verdict

Use Probe0 if your bill comes from coding agents on your own Mac. You get local model routing, caching that never touches the network, a cap that pauses instead of alerting, and a ledger that names the process that spent each dollar. Vercel AI Gateway is the better pick in one situation: you are routing production application traffic for other people and need a broad model catalog with automatic failover across providers. That is a different job. For the terminal you have open right now, Probe0 is the one that sees the traffic and can stop it.

Questions

Can I use Vercel AI Gateway with Claude Code or Cursor?
Not in the way people usually mean. The Gateway is designed for application code calling it over HTTPS, typically through the Vercel AI SDK, so it does not sit under an AI coding agent CLI by default. Probe0 works the other way around: a local proxy plus one certificate, and every agent CLI on the machine routes through it without per-tool config.
Is there a self-hosted or local Vercel AI Gateway alternative?
Vercel does not offer the Gateway as a self-hosted product; it always runs on their infrastructure. Probe0 is a local LLM proxy that runs entirely on your machine, with no Probe0 server involved, so cached prompts and locally routed requests never leave the laptop.
Does Probe0 support local models like Ollama and LM Studio?
Yes. Local Routing sends work to a model already running in Ollama or LM Studio, and automatically retries on a cloud model when the local answer comes back weak. A hosted gateway routes to hosted provider APIs, so local inference is outside its scope.
Which one gives better LLM cost control?
It depends on what you are controlling. Vercel AI Gateway is stronger for organisations, with budgets per key, project and team that return a hard refusal when exceeded. Probe0 is aimed at one developer: a hard cap per run and per day that warns then pauses, plus a ledger of model, tokens, real cost and the process that spent it, and advice on whether your subscription tier matches your actual usage.

Get started

Switch from Vercel AI Gateway to Probe0