ProofHound
proofhound.org
- Free plan
- Web
- API
- Self-hosted
What ProofHound is
ProofHound automates prompt optimization for LLM classification workflows. It analyzes error cases, iterates prompts, validates outcomes, and supports deployment and rollback. Users can upload CSV, TSV, JSONL, JSON array, or ZIP datasets and map fields in the interface. Optimization can target overall accuracy or category-specific measures such as recall for high-risk categories and precision for error-prone classes. Immutable prompt versions preserve variable configurations, output rules, and differences. Experiment records include datasets, model configurations, sample judgments, and overall and category metrics. Releases can use gray traffic, A/B testing, full rollout, and one-click rollback. Connections include the web UI, webhooks, API tokens, and MCP. Customers bring their own model provider, and ProofHound says it does not charge per model call. Free and self-hosted open-source options cost nothing. Pro is planned at $29 a month, billed Coming soon, with checkout not open yet.
Who it suits
ProofHound targets critical classification work in risk control, financial judgment, content moderation, and customer-service intent recognition. Its low-code positioning also suits operations, risk, and analyst teams.
Good to know, and what to watch
- Supports multiple dataset upload formats
- Category-specific optimization targets
- Immutable prompt version history
- Deployment includes rollback controls
- Self-hosted option supports private infrastructure
- Pro checkout is not open yet
- Free plan limits projects, members, calls, storage, and uploads
- Customers must bring their own model provider
- Hosted data remains until user deletion subject to quota
Plans
Every plan and priceFree
Self-hosted open source
Pro
Free plan
The verdict
ProofHound provides a structured path from classification error analysis through prompt release and rollback. Its free and self-hosted options are available now, while the planned Pro tier and upcoming generative-task features are not yet available.
What the maker publishes
- Dataset inputs
- The product supports CSV, TSV, JSONL, JSON array, and ZIP dataset uploads, with flexible field mapping in the UI. source
- Experiment audit
- The platform records prompt versions, datasets, model configurations, sample judgments, and overall and category metrics for traceable and reproducible experiments. source
- Integrations
- The site lists Web UI, Webhook, API Token, and MCP as connection options for business systems and AI agents. source
- Intended users
- The product targets critical classification workflows including risk control, financial judgment, content moderation, and customer service intent recognition, including low-code work by operations, risk, and analyst teams. source
- Model usage
- Customers bring their own model provider, and ProofHound says it does not charge per model call. source
- Optimization targets
- Users can optimize overall accuracy or tune category-specific metrics such as recall for high-risk categories and precision for error-prone classes. source
- Production controls
- ProofHound supports gray traffic release, A/B testing, full rollout, and one-click rollback for prompt deployments. source
- Purpose
- ProofHound automates prompt optimization for LLM classification tasks by analyzing error cases, iterating prompts, validating results, and supporting deployment and rollback. source
- Roadmap
- Evaluation, comparison, and optimization for generative LLM tasks and ProofHound Cloud Managed Enterprise Edition are listed as upcoming. source
- Security and data control
- The self-hosted edition supports private deployment and private data storage on the user's own infrastructure; hosted data is retained until user deletion subject to plan quota. source
- Support
- The site lists GitHub, Discord, a Chinese-speaking QQ group, and email for community discussion, product updates, and business contact. source
- Version history
- ProofHound records immutable prompt versions with variable configurations, output rules, and version differences. source
Checked against the maker's pages on 2 Oct 2026.