Probe0 vs Not Diamond
Not Diamond tells you which model to call. Probe0 sits in the path and makes the call, or avoids making it. One is a recommendation service you integrate into your code; the other is a local proxy your existing coding agent CLIs already route through.
A hosted intelligent model router: you send request context, it returns a recommendation for which model to use, and you make the actual LLM call yourself.
| Feature | Probe0 | Not Diamond | Edge |
|---|---|---|---|
| Where your request context goes | Stays on your Mac. No Probe0 server exists. | Sent to a hosted API. VPC only at enterprise tier. | Probe0 |
| Setup for one developer | One local proxy plus a certificate. Claude Code, Codex and Cursor all route through it. | SDK or REST call you add to your own code, per tool. | Probe0 |
| Response caching | Exact cache on local disk plus a local semantic index. | None. Cache state is a routing input, not a store. | Probe0 |
| Cache safety around tool calls | Semantic cache refuses to match anything carrying tool calls. | No cache, so nothing to protect. | Probe0 |
| Hard spend limits | Spend Guard: per-run and per-day cap. Warns, then pauses. | Cost is a tunable objective. No cap found in docs. | Probe0 |
| Cost attribution | Ledger of real calls: model, tokens, cost, latency, calling process. | Dashboard derived from routing events, not your actual traffic. | Probe0 |
| Behaviour with no network | Cache hits and local models keep working offline. | Every routed request needs a round trip to the API. | Probe0 |
| Savings you can attribute | Each module switches on alone and reports what it saved. Flags an over-sized plan tier. | Aggregate cost objective across routed requests. | Probe0 |
| Routing quality | Heuristics I wrote: tier order, local-first, retry weak answers. | Learned meta-model, trainable on your own eval data. | Competitor |
Not Diamond's router is a learned model trained on evaluation data, and you can train a custom one on your own prompts, candidate responses and eval scores, which is real machine learning infrastructure aimed at picking the better answer rather than the cheaper one. What it does not do is touch your inference call: the routing API hands back a session ID and a model name, and you make the request yourself with your own SDK and keys. For a developer running coding agents on a laptop, that boundary is where the money leaks. A recommendation cannot serve a cached response, cannot stop a request at a budget threshold, and cannot tell you which process on your machine spent the money. Probe0 sits in the path, so it can do all three.
The integration cost differs in the same direction. Not Diamond is a call you add to your own code, which is fine when you own the code, and useless when the caller is Claude Code or Codex or Cursor and you own none of it. Probe0 is one local proxy and one certificate installed once; after that every coding agent CLI on the machine routes through it with no per-tool configuration and no code change. Nothing is sent to me for a routing decision, because there is no me to send it to. Each Not Diamond routed request, by contrast, costs a network round trip on top of your inference bill, and its savings dashboard is assembled from routing events rather than from the calls that actually happened.
I built Probe0 for the case where you are one developer on one Mac with three coding agent CLIs open and a bill you do not fully understand. Exact Cache serves repeats from local disk without touching the network. Semantic Cache uses a local vector index with a strict similarity floor, and refuses to match anything carrying tool calls, because a wrong hit there corrupts a run rather than just annoying you. Request Coalescing collapses simultaneous identical calls into one upstream. Local Routing sends work to a model already loaded in Ollama or LM Studio and retries weak answers on the cloud. Spend Guard actually pauses at your cap. Every module is switchable on its own and reports what it saved, and because the recording ledger sees real calls with per-process attribution, it can also tell you when your plan tier is above what you use. Honest limits: macOS only, private beta, sign-in through Google or GitHub, no multi-user gateway, and provider coverage is whatever coding agents call rather than a hundred-provider catalogue.
The verdict
Pick Not Diamond if you are shipping a product whose output quality depends on model choice and you have eval data to train a router on. For everyone else running coding agents on a Mac, Probe0 is the one to install: it needs no code change, it eliminates calls instead of relabelling them, it shows you which process spent what, and its cap stops the run instead of sending an alert after the money is gone.
Questions
- Is Not Diamond a proxy or a gateway?
- Neither. It returns a routing recommendation and you make the LLM call yourself with your own SDK and keys. Its own docs position it as complementary to a gateway rather than a replacement for one.
- Does Not Diamond have spend caps or caching?
- No spend caps or hard budget enforcement appear in its docs or pricing, and there is no response cache. Cost is a tunable routing objective, and cache context is an input to routing rather than a store of answers.
- What is the best Not Diamond alternative for a solo developer?
- If you want a local LLM proxy rather than a hosted router, Probe0 runs entirely on your Mac and covers Claude Code, Codex and Cursor with one proxy and one certificate. It handles local LLM routing, caching and hard spend limits without shipping request context anywhere.
- Can Not Diamond route to a local model in Ollama?
- It can recommend a custom model endpoint you register, but the router must be trained on eval data for that model first, and Not Diamond never calls the endpoint. Probe0 sends the request to a model already running in Ollama or LM Studio and retries weak answers on the cloud.