In a large organisation, AI adoption does not fail on model quality. It fails on distribution: the practices that work stay with the handful of people who worked them out. Grimoire makes those practices a package.
The problem
One company runs Claude Code, Copilot, Cursor, Codex, Gemini, Zed, Amp, Kiro, Junie and opencode side by side, none of them reading the same configuration format. A team that works out a good rule or a well-scoped agent can only hand it on as a document someone is meant to copy by hand — so the enthusiasts pull ahead, everyone else never starts, and nobody can say which guidance is actually running.
What I built
grim takes a declaration — skills, rules, agents, MCP servers — and writes each one
into every client in the format that client actually reads, pinned by digest in a
lockfile. One person writes the practice; everyone else runs grim install and has
it in the agent they already use.
Storage is any OCI registry, including an internal one, so the artifacts inherit the authentication, mirroring and audit trail the organisation already trusts. Nothing has to leave the network and there is no Grimoire service to sign up for.
Where a client has no honest surface for an artifact, grim says so and skips it
rather than writing configuration that looks installed and quietly does nothing. A
package manager that lies about what it installed is worse than none.
Where it stands
Apache-2.0, stabilizing toward 1.0 — released surfaces are frozen contracts. Pre-built binaries for macOS, Linux and Windows on both architectures, a one-line installer, a self-hostable search index, and a GitHub Action for CI.