GEPAzilla
gepazilla.com
- Free plan
- Web
- API
- Self-hosted
What GEPAzilla is
GEPAzilla is an open-source prompt optimizer for improving system prompts with datasets and scorers. It can run as a web app, API, or self-hosted Next.js application, and the project is released under the MIT License. You can import your own dataset or use its sample, though dataset tools currently accept JSON payloads copied or pasted from the clipboard. Automated optimization combines deterministic checks with LLM judges, configurable scorer weights, and duplicate handling. Validation rows remain outside reflection and help rank candidate prompts on a Pareto frontier. Results include candidate prompts, iteration trends, scorer failure rates, averages, and common notes. GEPAzilla also optimizes request latency, with a latency metric always shown in Results. GEPA runs, scorer previews, and model discovery require a Vercel AI Gateway API key. The maker says run telemetry, including tokens, routing, costs, and errors, stays on your machine; events are not sent remotely and run data is not exfiltrated. The software is free and open source.
Who it suits
It suits curious builders experimenting with prompt optimization using their own API key. It also fits users who want web, API, or self-hosted deployment options.
Good to know, and what to watch
- Open-source MIT-licensed project
- Web, API, and self-hosted options
- Combines deterministic checks and LLM judges
- Validation rows rank candidate prompts
- Latency metric included in Results
- Requires a Vercel AI Gateway API key
- Dataset import is clipboard JSON only
The verdict
GEPAzilla provides automated prompt testing with configurable scoring, validation separation, and diagnostic results. Its API-key requirement and clipboard-only dataset import are important setup considerations.
What the maker publishes
- API key requirement
- The project README says an AI Gateway API key is required for GEPA runs, scorer previews, and model discovery. source
- Built-in latency metric
- GEPA optimizes request latency internally, and the Results tab always includes a latency metric. source
- Data handling
- The README says telemetry events are not sent to a remote service and run data is not exfiltrated. source
- Data handling
- The README says telemetry stays in the browser, telemetry events are not sent to a remote service, and run data is not exfiltrated. source
- Data import limit
- Dataset tools currently support copying or pasting JSON payloads from the clipboard. source
- Dataset format limit
- Dataset tools currently support copying or pasting JSON payloads from the clipboard. source
- Dataset input
- You can load your own dataset or start with the included sample dataset. source
- Dataset input
- Users can load their own rows or start with the sample dataset. source
- Deployment
- The project is a standard Next.js application that can be built and started with pnpm. source
- Deployment
- The README describes the application as a standard Next.js project that can be built and started locally. source
- Diagnostics
- Results include scorer failure rates, averages, and common notes, alongside candidate prompts and iteration trends. source
- Gateway integration
- The README specifies a Vercel AI Gateway credential for the required API key. source
- Intended users
- The maker describes the console as a way for curious builders to experiment with prompt optimization using their own API key. source
- License
- The project is released under the MIT License. source
- License
- The repository states that GEPAzilla is released under the MIT License. source
- Model access
- The project requires users to supply an AI Gateway API key for GEPA runs, scorer previews, and model discovery; its README specifies Vercel AI Gateway. source
- Purpose
- GEPAzilla is an open-source GEPA prompt optimizer that iterates on system prompts using datasets and scorers. source
- Purpose
- GEPAzilla is an open-source GEPA prompt optimizer that uses datasets and custom metrics to find stronger prompts. source
- Scoring
- It supports combining deterministic checks with LLM judges, with configurable scorer weights and duplicates. source
- Scoring
- It supports combining deterministic checks with LLM judges and weighting or duplicating scorers. source
- Telemetry
- The maker says run telemetry stays on your machine and includes tokens, routing, costs, and errors. source
- Telemetry
- The maker says telemetry stays local and includes tokens, routing, costs, and errors for inspection. source
- Validation
- Validation rows are kept out of reflection and used to rank candidate prompts on the Pareto frontier. source
- Validation data
- Validation rows are kept untouched during reflection and used to rank candidates on the Pareto frontier. source
Checked against the maker's pages on 10 Oct 2026.