The documents live in the repository as Markdown and travel with it, so a decision recorded from one coding agent is waiting for the next, and edits to it get reviewed in a pull request like any other file. The tools stay deliberately plain — list, get, search, create, update, link — which leaves the structure of the directory up to how your team already thinks about the codebase.
A local stdio server shipped inside the Archcore CLI. It keeps typed Markdown documents — decision records, rules, specs, plans, incident write-ups — in a directory inside your repository, and lets any MCP-compatible coding agent read, write and link them while it works.
- Documents an agent can create, read, update and remove, each typed by what it is rather than filed as loose notes
- Search across them, so an agent loads the decisions covering the area it is about to touch and nothing else
- Directed links between documents: one implements another, extends it, depends on it, or is merely related
- Bootstrapping in a repository that has no context yet — the agent sets the directory up itself
- Prompts that chain a whole cascade in one call, an idea through a requirements document to a plan
- Everything on disk as Markdown, versioned in Git and reviewed in pull requests like the code beside it
No account and no API key — the server reads and writes files inside your repository. It is a Go binary you install first, and the CLI then writes the MCP configuration for the coding agents it detects. Because it serves the directory it was started in, an agent launched from elsewhere has to be told which repository to serve.
One command — curl -fsSL https://archcore.ai/install.sh | bash