Skip to content

Authoring

You don’t author a sightmap by hand.

A sightmap describes how a web app behaves at runtime — its routes, components, network calls, and the institutional knowledge an agent needs to drive it. .sightmap/ is curated authority, not a generated artifact: the agent observes the running app, proposes the YAML, and commits it. The tools in this section are the curation surface.

Curation runs in two phases:

  1. Bootstrap — once, when you first add Sightmap to a repo. sightmap init scaffolds an empty .sightmap/app.yaml, wraps your app entry with <SightmapProvider>, and wires the MCP server into your agent harness. You add the Vite plugin, start your dev server, and an agent walks the running app: it fetches the live fiber-tree, proposes views and components with ranked selectors, and commits the batch to .sightmap/.
  2. Maintain — every time routes change, components move, or the agent learns something new about the app, you re-run the bootstrap loop or call the curation tools directly. Existing fields (memory, requests, manual selector overrides) are preserved.

The agent owns the entire corpus. There is no source-side codegen; static drift detection was removed in v0.6. Reconciliation is agent-driven, against the runtime.

Different surfaces handle different parts of the loop:

  • sightmap init (@sightmap/sightmap) scaffolds the empty corpus, codemods the provider, and wires the agent harness. One-shot, idempotent.
  • The React runtime (@sightmap/react) exposes the live fiber-tree at /__sightmap__/snapshot.json. It never writes to .sightmap/.
  • An agent (via the MCP server, the Claude Code plugin, or any host) writes the corpus — proposing from the runtime snapshot, committing accepted proposals, patching fields directly with sightmap_update_view.
  • You rarely touch YAML directly. When you do, sightmap check validates the corpus against itself — schema, key ordering, cross-file refs, duplicates.
  • @sightmap/react — the React runtime. Mount the provider, add the Vite plugin, and an agent bootstraps .sightmap/ against the running app.
  • @sightmap/cli — the sightmap command-line tool. validate, lint, match, explain. Used in CI and during day-to-day curation.
  • @sightmap/core — the library API behind the CLI. For tool builders integrating sightmap into IDEs, MCP servers, or custom workflows.
  • @sightmap/mcp — an MCP server combining browser introspection with .sightmap/ curation tools. The recommended way to let a host agent author for you.
  • @sightmap/playwright — sightmap-aware wrapper around @playwright/cli. Same enriched snapshots as @sightmap/mcp over a shell-invocation surface — the no-MCP path.
  • @sightmap/agent-browser — sister wrapper for Vercel’s agent-browser CLI. Same four enriched commands; passes through -p <provider> for hosted Chrome (Browserless, Browserbase, Kernel, …).
  • Plugin (Claude Code) — advisory hooks that nudge Claude Code toward MCP curation tools and validate every .sightmap/ mutation.

If you’re starting from scratch, follow the quickstart — it walks the bootstrap-and-maintain loop end to end.