Introducing ProofHound Cloud Beta

ProofHound began as an open-source tool for one stubborn problem: tuning the prompts behind LLM classification — the models that sit in risk control, content moderation, financial review, and customer-intent routing. That work is mostly invisible and mostly manual. Someone reads through the cases the model got wrong, rewrites the prompt, runs it again, compares the numbers, and repeats. The judgment that matters is a thin slice of the effort; the rest is mechanical, rarely written down, and slow to improve.

The open-source edition automates that loop. But to run it, you first stand up a database, a queue, object storage, and workers — and then keep them alive. For a team that simply wants to find out whether automated prompt optimization helps, that is a lot of groundwork before the first experiment.

ProofHound Cloud Beta removes it. You sign up, create a workspace, and start. There is nothing to deploy and nothing to maintain.

A hosted home for the whole team

The hosted version is built around organizations rather than individuals. An organization is where your projects live, who can see them, and how usage is counted. Invite teammates and give them roles from owner down to viewer, and they share the same projects, the same history, and the same quota — instead of each person running a private copy.

Inside a workspace you get the full ProofHound product, unchanged. Import a labeled dataset, watch the optimizer analyze the failures and rewrite the prompt, compare runs side by side, and promote a winning version through a controlled release with the option to roll back. The console for models, datasets, prompts, experiments, optimizations, connectors, releases, and annotations is the same one open-source users already know.

Connecting your own systems works the same way it does in open source. A personal access token reaches your projects over the HTTP API or through MCP, so an agent or a backend service can drive the same workflows you run in the browser — and the token never grants more than your own role allows.

We are opening with a Free plan: three projects, one member, three concurrent model calls, and 5 GB of retained storage, with a heads-up as you approach each limit. You bring your own model provider, and ProofHound does not charge per call or meter your samples and runs. The plan is about workspace capacity, not usage tolls.

Who it is for

ProofHound Cloud is aimed at teams whose classification decisions carry weight — flagging risk, judging a transaction, moderating content, routing an intent — and especially at the awkward cases where one small but high-stakes category cannot be allowed to disappear into an average. Operations, risk, and analyst teams can configure rules and iterate prompts directly in the interface without writing code, and try all of it on a managed workspace before deciding whether to self-host.

What has changed in open source since launch

The hosted version rests on an open-source project that has been moving quickly. Since the first versioned release at the end of May, most of the work has gone into the parts you feel when a dataset gets large. Imports now stream in the background instead of freezing the page, run with automatic concurrency, map fields visually, accept everything from CSV to zipped archives, and stop cleanly if you leave. Large run results and datasets now spill over into object storage instead of bloating the database — the same foundation the hosted version uses for its own storage.

Underneath, runs have grown more reliable. Stopping a run now also cancels the model calls still waiting in its queue, results are kept month by month, and result handling was hardened against partial failures. A quieter but important thread of work finished the seams that let the hosted version reuse the open-source code without ever forking it, so quota policy, organization-aware rate limiting, and usage accounting all attach cleanly. And a run of smaller refinements — time-zone-aware timestamps, calmer loading states, automatic language detection — left the everyday surface feeling steadier.

What is next

Two things are close. In mid-July we will open Pro, a paid plan built for teams rather than a single seat.

Pro lifts the ceilings the Free plan sets: projects and members with no hard cap, all sharing one organization quota; fifty concurrent model calls; workflows that can run for up to seven days; and 50 GB of retained storage. It includes full role-based access control and every integration channel — access tokens, connectors, webhooks, and MCP. The point is not just bigger numbers. It is that a team can work inside one organization, share assets and quota, and pay together, instead of stitching separate environments into something that resembles collaboration. We are planning $29 a month, and you can reserve a seat from the pricing page to hear first when it opens.

Beyond Pro, the work is steady rather than flashy. We will keep smoothing the everyday interactions in the interface, and keep improving the optimization strategy itself so prompts converge faster and waste fewer dead-end attempts. Further out, we are extending ProofHound past text classification: support for the video modality, and support for generative tasks — bringing the same evaluate, compare, and optimize loop to problems that do not have a single correct label.