GEPA

GEPA is an open-source engine for optimizing prompts and other text-representable material, including code, configurations, agent architectures, and policies. It uses execution traces and evaluator feedback to guide LLM reflection, propose targeted changes, and select candidates across a Pareto frontier. Its optimize_anything API can connect to a system through a user-written evaluator without requiring a particular framework. GEPA works with API-only models such as GPT, Claude, and Gemini without access to model weights. Integrations include DSPy, MLflow, Comet ML Opik, Pydantic AI, OpenAI Cookbook, Hugging Face Cookbook, and Google ADK. Built-in adapters cover workflows such as RAG, MCP, TerminalBench, and LangChain. Human-readable traces show why changes were made and can aid debugging. Users can cap metric calls and set multiple stopping conditions. The project says improvements can appear with as few as three examples and recommends 30–300 for best results. GEPA is a free, self-hosted Python package installed with pip and listed under the MIT license. It does not inherently optimize for short prompts; prompt length can be added as an objective.

GEPA offers a flexible, trace-guided approach to optimizing prompts and other text-based artifacts, with adapters for several tools and frameworks. Its prompt-length limitation and reliance on evaluator feedback are worth considering when setting an optimization objective.

GEPA may suit teams optimizing expensive rollouts, working with scarce examples or API-only models, or needing readable traces of prompt changes. It is also relevant when a user-written evaluator can represent the system's objective.

GEPA is an open-source Python framework for improving prompts and other text-based parts of AI systems against evaluation metrics. It suits teams with costly rollouts, limited examples, or models available only through APIs. Its trace-guided approach makes changes inspectable, though teams must include prompt length as an objective if concise prompts matter.

Overview

GEPA uses execution traces and evaluator feedback to diagnose failures, guide targeted changes through LLM reflection, and select candidates with a Pareto frontier. Evaluators can supply error messages, profiling data, and reasoning logs, so the process can work with diagnostic signals beyond a single score. That makes GEPA useful for improving prompts, code, configurations, agent architectures, policies, and other text-representable artifacts. It is a framework for optimization, rather than a ready-made prompt editor.

The optimize_anything API works with any system for which a user can write an evaluator, without requiring DSPy or another framework. GEPA can optimize API-only models such as GPT, Claude, and Gemini without access to their weights. Its human-readable traces help teams understand why a prompt changed, which is useful when debugging agent behavior matters alongside improving evaluation results.

Key features

• Trace-guided changes: LLM reflection uses execution traces to propose targeted mutations. This can make expensive or scarce rollouts more informative, especially when evaluators provide useful diagnostics. The quality of those signals matters: GEPA's direction depends on the evaluator's feedback.
• Flexible evaluation: The user-written evaluator in optimize_anything lets teams apply GEPA to systems outside a supported framework. That flexibility is valuable for custom stacks, but it also means teams need to define and maintain the evaluation logic themselves.
• Framework adapters: Built-in options cover Default, Confidence, DSPy Full Program, Generic RAG, MCP, TerminalBench, AnyMaths, and LangChain. The project also lists integrations with DSPy, MLflow, Comet ML Opik, Pydantic AI, the OpenAI and Hugging Face Cookbooks, and Google ADK. This gives teams several starting points without removing the option to write a custom evaluator.
• Budget and stopping controls: Users can cap metric calls and set timeout, no-improvement, score-threshold, signal, file, and composite stop conditions. Those controls help bound an optimization run and end it when progress or a chosen target justifies stopping.
• Small-data operation: GEPA can show improvements with as few as three examples; the project recommends 30–300 for best results. It also describes 100–500 evaluations as a use case for expensive rollouts. These figures make it relevant when data or evaluation opportunities are limited, while abundant data and 100,000 or more cheap rollouts can favor gradient-based methods.
• Agent skill: The repository includes a GEPA Agent Skill for coding agents including Claude Code, Cursor, VS Code/Copilot, Codex, and Gemini CLI. Teams using those tools can incorporate the project through that skill as well as its Python package.

GEPA's main trade-off is prompt length. It does not inherently seek shorter prompts, so optimization can produce longer ones unless length is part of the objective. That is a meaningful constraint for teams managing context budgets or prioritizing compact instructions.

Pricing

GEPA is free and open source under the MIT license. The GEPA plan costs nothing and provides an open-source Python package installed with pip install gepa. There is no paid tier or seat quota described in the plan. This makes it a straightforward fit for teams comfortable running a Python package and bringing their own evaluation setup; its cost advantage does not remove the work of creating that setup.

Platforms

GEPA is available through an API and as self-hosted software. Its Python package and framework adapters suit teams that can integrate optimization into their own systems. The project also provides a documented path for coding-agent use through its Agent Skill.

Who it's for

GEPA is best suited to teams that need to improve AI behavior with expensive rollouts, scarce examples, or API-only models, and that want human-readable traces to inspect changes. It is also a good candidate for custom systems whose owners can write an evaluator. Teams with ample data and at least 100,000 inexpensive rollouts may find gradient-based approaches more effective, while teams that need shorter prompts should explicitly include prompt length in their optimization objective.

Pros and cons

• Pro: Evaluator diagnostics can direct targeted changes, making limited or costly evaluations more useful than score-only feedback.
• Pro: Human-readable traces show why prompts changed and can support agent debugging.
• Pro: The custom-evaluator API and built-in adapters accommodate both framework-based and bespoke systems.
• Pro: The MIT-licensed package is free, and API-only model support avoids requiring access to model weights.
• Con: Teams must supply an evaluator, and the optimization depends on the quality of its feedback.
• Con: Prompt length is not optimized by default, so concise outputs require an explicit length objective.

Alternatives

AI Prompt Generators is a useful directory to consult when comparing prompt-generation tools across the category.

• Vellum is a better fit for readers seeking a freemium option across Android, iOS, desktop, web, browser extension, and self-hosted platforms, with a free tier that includes initial tasks and pay-as-you-go credits.
• Promptitude suits readers who want a web and API prompt platform with a free tier capped at two users, five prompts, two assistants, one flow, one tool, 100 generations per month, and 200 tool calls per month.
• PromptWizard is another free, open-source option for API, self-hosted, Linux, macOS, and Windows use.
• VantagePrompt is worth considering for a web-based freemium tool with a free allowance of 25 monthly credits and defined request limits.
• Agenta may suit teams looking for a web, API, or self-hosted evaluation platform; its free Hobby plan includes two team members, 5,000 agent runs per month, 20 evaluations per month, and one-week trace retention.
• ChainForge is a free, open-source choice for web or self-hosted use.
• Google Cloud Agent Evaluation is an option for readers who prefer a paid, usage-priced service with API and web access and a free trial.
• Latitude offers a freemium alternative for API, self-hosted, and web deployments; its free Starter plan includes 20K credits per month, 30-day data retention, unlimited seats, and general-channel Slack support.

Verdict

Choose GEPA if you can provide a meaningful evaluator and want a free, inspectable way to optimize prompts or other text-based AI components, particularly when rollouts are costly or examples are scarce. Its broad adapter set and API-only model support make it adaptable, while evaluator setup and the need to specify prompt length as an objective are the main reasons to look elsewhere.