Prototype 0 is a local Go CLI backed by a .glyph/ directory. There is no hosted Glyph service yet and no long-running daemon required. Humans and agents invoke the glyph CLI when they want to read, write, checkpoint, publish, visualize, or export source.

Core Concepts In One Diagram

If Git is a commit graph plus conventions, Glyph is a source graph with these concepts built in.

Local Architecture

store.db stores graph structure and metadata. .glyph/content/ stores file bytes by content identity so repeated file states can share physical storage. The system stores full file states logically, but deduplicates bytes physically.

What Lives Where

PathRole
.glyph/store.dbSQLite database for graph metadata, realms, work contexts, snapshots, claims, publications, remotes, and mounts.
.glyph/content/Content-addressed file bytes.
.glyph/workspaces/Materialized workspace projections for active work.
.glyph/hooks/Local executable hooks.
.glyph/exports/Generated compatibility exports, such as clean Git repositories.
glyph.yamlProject policy and defaults for imports, exports, and visualizer output.

Read And Write Path

The CLI is the control point. An agent does not need shell access to Git internals; it can use glyph read and glyph write as the source-control API.

Publication Path

Publication changes a realm. Remote sync exports that realm to GitHub. Those are separate steps so Glyph can keep richer local semantics while still using GitHub infrastructure.

Does The Agent Run Glyph On Its Own?

Glyph does not autonomously decide to run. Something invokes it:
  • a human in a terminal
  • a coding agent following project instructions
  • a script or CI job
  • a future hosted Glyph service
In prototype 0, the recommended agent integration is simple: give the agent a skill or instruction document that says to use glyph --json, claim work before editing, checkpoint progress, check conflicts, publish intentionally, and sync remotes only when asked. This keeps the integration portable across Codex, Claude Code, Cursor, and other agents. A future MCP server or hosted API can provide richer integration, but the CLI is enough to dogfood the model today.