potpie-ai/potpie

Potpie is a platform for building task-oriented AI agents that operate directly on a codebase, aimed at engineering tasks like debugging, integration testing, system design, and developer onboarding rather than general-purpose chat. Its agents are built to be codebase-aware, drawing on a knowledge graph of the code along with logs, pull requests, and existing workflows rather than working from prompts alone, which the company positions as the difference between "surface-level AI" and agents that hold up at large scale. Potpie reports scoring 63% on the SWE-bench Lite benchmark, is open source with over 5,100 GitHub stars, and says it has been used on codebases exceeding 50 million lines across Fortune 500 and regulated-industry engineering teams. Teams can create custom agents for specific workflows — from migrations to refactors to architecture reviews — and manage them through Potpie's build system.