Probe0 vs Martian
Martian is a hosted router that picks a cheaper model for you, and it does that with real published research behind it. Probe0 is a local proxy that sits on your Mac in front of Claude Code, Codex, and Cursor, and cuts spend by not sending the request at all when it does not have to.
A hosted, OpenAI-compatible LLM gateway with a trained model router that predicts which model can answer a given prompt acceptably and sends it to the cheapest one.
| Feature | Probe0 | Martian | Edge |
|---|---|---|---|
| Where your prompts go | Stay on your Mac, no Probe0 server exists | Every request transits their hosted endpoint | Probe0 |
| Routing intelligence | Rules you configure: cheap model first, local model first | Trained router predicting per-prompt quality, ROUTERBENCH authors | Competitor |
| Setup for one developer | One proxy plus one certificate, all three agent CLIs covered | Point each SDK or tool at their base URL | Probe0 |
| Models running on your own machine | Ollama and LM Studio first, weak answers retried on the cloud | No evidence of local model routing | Probe0 |
| Caching | Exact disk cache plus local semantic cache, zero-network hits | Not a documented feature | Probe0 |
| Tool-call safety in cache | Semantic cache refuses to match anything carrying tool calls | No cache, so no policy | Probe0 |
| Spend limits | Hard cap per run and per day: warns, then pauses | Prepaid credits, no documented per-key budgets | Probe0 |
| Cost attribution | Local ledger: model, tokens, real cost, latency, per process | Hosted dashboard, aggregate spend on the key | Probe0 |
| When a provider has an incident | Falls back to a local model, keeps working offline | Reroutes across providers | Even |
Martian's router is the one thing here I would not try to beat: a trained model that predicts how well each candidate will answer your specific prompt before running any of them, distilled from input and output pairs, published as ROUTERBENCH, funded by Accenture to take that routing to clients. Everything else about the comparison points the other way for a developer running coding agents on a Mac. Probe0 is an LLM proxy that lives locally, so proprietary code stays on the machine unless a module decides a cloud call is warranted, and there is no Probe0 server anywhere for prompts to transit. One proxy and one certificate cover Claude Code, Codex, and Cursor together, with no per-tool configuration after that. A hosted gateway needs each SDK repointed at its base URL, and every request it handles crosses the network by definition.
The savings come from requests that never leave. Local routing sends work to a model already loaded in Ollama or LM Studio and quietly retries on the cloud when the answer is weak. Model tiering tries the cheap model first and escalates only when needed. Exact cache hits read from local disk and touch no network at all, which also means they work on a plane. The semantic cache runs a local vector index behind a strict similarity floor and refuses to match anything carrying tool calls, because a wrong hit there corrupts an agent run in a way that costs far more than it saved. Request coalescing collapses simultaneous identical calls into one upstream. Every module switches on and off independently and reports what it saved, so you can judge each one on its own numbers instead of a single blended discount.
The other difference is guardrail versus optimizer. Martian optimizes cost by picking a cheaper model, and you pay a metered service on top of model rates for that mechanism. Prepaid credits are a ceiling in the sense that you run out, but public docs do not describe per-key spend rules. Probe0 does LLM cost control the blunt way: a hard cap per run and per day that warns and then pauses the agent, plus a full local ledger recording model, tokens, real cost, latency, and which process made each call. Per-process attribution is what tells you that Cursor, not Claude Code, burned the afternoon. That same ledger powers plan-tier advice, since knowing your actual usage is what lets me tell you when you are paying for a tier above what you use. One continuity note worth stating plainly: withmartian.com today reads mostly as an interpretability research lab, the Gateway sits under docs as one product among research outputs, and marketing subdomains from the router era no longer resolve. That is a fair question to ask before building on it, not an accusation.
The verdict
Pick Martian if cross-model quality prediction is the specific thing you want, because their trained router is the real research and my rules are just rules. For everything else a solo developer on macOS actually hits, I would run Probe0: prompts that stay on your machine, one certificate that covers Claude Code, Codex, and Cursor at once, cache hits and local models that never touch the network, a cap that pauses the agent instead of emailing you about it, and a ledger that names the process that spent the money. Probe0 is macOS only and in private beta. It is still the one I would install first.
Questions
- Is Probe0 a Martian alternative?
- For a solo developer running AI coding agents, yes. It is a different shape though: Martian is a hosted router that picks a cheaper model per prompt, while Probe0 is a local LLM proxy that cuts spend through caching, local models, coalescing, and a hard spend cap. Probe0 has no equivalent to Martian's trained quality-prediction router.
- Do my prompts leave my machine with Probe0?
- Only when a request actually needs a cloud model. There is no Probe0 server, so nothing is relayed through us. Cache hits and local LLM routing to Ollama or LM Studio never touch the network at all. With Martian, every request transits their hosted infrastructure by design.
- Can Martian route to a model running on my laptop?
- There is no evidence it can. Martian is a hosted cloud gateway routing to hosted frontier providers, so a model on your own machine is out of scope architecturally. Probe0 treats a locally running model as the first thing to try and retries on the cloud when the answer looks weak.
- Which one gives better LLM cost control for a hard budget?
- Probe0, if your goal is a stop rather than a discount. It enforces a cap per run and per day, warning first and then pausing, and records every call with real cost and the process that made it. Martian bills against prepaid credits and does not publicly document per-key budget rules, though its router is the stronger cost optimizer.