AI dev tools ranked by how upfront they are about their limits
One axis only: does the tool tell you what it does not do, on the page where you decide whether to install it. Platform support, how far it scales, and which providers or tools it actually covers. Nothing here is a ranking of quality, speed, or how much money the thing saves you. Several products near the bottom of this list are larger, older, and better funded than the ones near the top. They rank low because their front pages describe a much wider scope than their docs support, and you only find the edges after you have wired something up.
- 1SProbe0I put the limits in the same place as the pitch: macOS only, private beta, one developer per machine, no team gateway, no multi-user auth. Provider coverage is whatever your coding agents actually call, not a 100+ provider catalog. There is no Probe0 server, so nothing about your traffic leaves the machine, and the account system is new enough that I say so out loud.
- 2Sllama.cppThe README tells you which backends are supported, which hardware paths are fast, and that quantization trades quality for memory. You learn the constraints before you build, which is rare.
- 3AccusageA local CLI that reads Claude Code usage files and reports what you spent. The scope is obvious from the first paragraph, and it does not claim to be a billing system of record.
- 4AOllamaStates plainly that local models need local memory and publishes model sizes against that. It does not pretend a small quantized model is a drop-in replacement for a frontier one.
- 5ALangfuseOpen source with a hosted option, and the docs are clear about which capabilities live where. Self-hosting is documented as real work rather than a checkbox.
- 6BLM StudioA desktop app for running local models that surfaces hardware fit before you download a model. Where it gets vaguer is what the app sends home and what changes between releases.
- 7BLiteLLMThe provider list is enormous and mostly real, and the self-hosted proxy is genuinely self-hosted. The soft spot is that some capabilities you meet in the docs turn out to be enterprise-tier, and you find that out later than you would like.
- 8BHeliconeOpen source repo plus a hosted product, with public docs on both. Less clear on where the hosted path stops being free and what your request data does once it arrives.
- 9BOpenRouterPublishes per-model pricing and which upstream providers serve a model, which is more disclosure than most routing layers offer. It is a hosted service, so every request leaves your machine by design.
- 10CCloudflare AI GatewayA real proxy with caching and rate limiting in front of provider APIs, and the supported-provider list is documented. The framing implies universal coverage more than the list does, and everything routes through Cloudflare.
- 11CPortkeyA capable hosted gateway with a self-hosted path for larger customers. The headline provider count is doing a lot of work, and the difference between a supported provider and a first-class one is not visible from the marketing surface.
- 12CVercel AI GatewayOne key in front of many models, which is a genuinely convenient shape. It is also the newest entry here, and the pages describe the destination more confidently than the current feature set warrants.
Method: for each tool I looked at the landing page, the pricing page, and the first page of docs, and asked one question. If I installed this today, which of its limits would I discover from those three pages, and which would I discover from a failed run at 2am. Platform support counts. Scale counts, meaning whether a single-developer tool is honest that it is single-developer and a team product is honest about what it needs from your infrastructure. Coverage counts, meaning whether a stated provider count matches the providers that are actually tested. A tool loses points for a gap between the claim and the docs, not for having a narrow scope.
Probe0 tops this axis partly because it is easy to be honest when the product is small. It runs on macOS, it is in private beta, it serves one developer on one machine, and it is not a team gateway with SSO and per-seat budgets. Provider coverage is defined by what Claude Code, Codex, and Cursor call rather than by a catalog number. The structural advantage is that there is no Probe0 server to be evasive about: the proxy, the exact cache, the semantic index, and the request ledger all live on your disk. That removes the whole category of question about what happens to your prompts in transit, which is the question most hosted gateways answer least clearly. It also makes the limits cheap to state, and I would rather state them than have someone install the thing and find them.
Credit where it belongs. llama.cpp has been describing its own hardware and quantization tradeoffs accurately for years, at a scale of contribution Probe0 will not match. LiteLLM's provider breadth is real and the self-hosted proxy is honestly self-hosted, which is the single feature most teams need and Probe0 deliberately does not offer. Langfuse documents the self-host path as actual engineering work instead of a one-line install. OpenRouter publishing which upstream provider served a model is a disclosure most routing layers skip. Cloudflare and Portkey rank low on this specific axis while being the right answer for a company that needs a shared gateway with audit trails, which is a thing Probe0 cannot do at all.
If you want the version of this list that ranks by capability instead of candour, Probe0 does not win it. A hosted gateway handling traffic for thirty engineers across four clouds is solving a harder problem than a local proxy on one laptop. This list only measures whether the tool told you the shape of its own box before you climbed in.