MCP

Run glyph mcp to start the stdio MCP server.

The server exposes glyph_explain, glyph_docs, glyph_doctor, glyph_list, glyph_spec_verify, glyph_spec_scaffold, glyph_run, glyph_snapshot_update, glyph_diff, glyph_context, glyph_render, glyph_repair, glyph_affected_specs, and glyph_clean. Tools call the same internal paths as the CLI so agents get the same validation, artifact packs, and exit behavior.

glyph_doctor runs the full prerequisite matrix (platform/PTY/config/artifacts/emulator), not a config smoke test.

glyph_list returns specs under given paths (name, path, coversSymbol).

glyph_spec_scaffold accepts kind: "spec" or kind: "action" so agents can create reusable action snippets without guessing the YAML shape.

glyph_render returns a deterministic SVG of a run's final screen (or a named snapshot). glyph_repair analyzes a spec's failed run and proposes step fixes; with write: true it applies them, only ever editing steps so the contract hash stays valid. Set verify: true for a transactional cold-start verification (SPEC §7.2) before applying.

glyph_affected_specs selects the specs a git change can hit: it shells out to codemap review --json and intersects each spec's coversSymbol against the changed symbols plus their blast radius, returning the minimal spec set to run via glyph_run. One of since/staged selects the diff scope; passing neither means the working tree.

glyph_clean prunes old run directories from the artifact root per retention.keepRuns (default 3), archiving pruned dirs to the configured external command before deleting them when retention.archive is set. all: true wipes every run dir; noArchive: true deletes locally without archiving.

Transport framing

The server speaks newline-delimited JSON-RPC over stdio (each message is a single JSON object followed by \n), which is what the MCP spec defines for stdio transport and what Claude Code, Codex, OpenCode, and Claude Desktop expect. The input reader also accepts Content-Length-framed requests for backwards compatibility with LSP-style clients, but responses are always line-delimited. Do not add Content-Length framing to the output — it breaks the Claude Code health check.