Metatron
Captures a team's real engineering conventions, rejected approaches and edge cases as versioned Markdown decision files in a repository, so a coding agent consults them before writing code.
Metatron records a codebase's actual implementation decisions, preferred patterns, rejected approaches and internal conventions as one Markdown file per convention, versioned in git alongside the code and curated through ordinary pull-request review. Any coding agent that can read a file can consult these decisions before writing code, and record what it learns as it goes, aiming to make an agent write code like an engineer who already knows the codebase.
It installs with pip and a single setup command that wires a repository to carry its own agent context as plain files, with no server required. Teams that want a serving layer on top can also run Metatron as a self-hosted MCP server backed by SQLite, which relevance-ranks results and takes an agent feedback loop, while files and server stay losslessly in sync. Metatron is a reference implementation of a proposed standard for git-native, agent-maintained project context.