@feelreef/mcp

Edit a feelreef native site from any MCP client — Claude, ChatGPT, your own agent. Every edit lands on the site's own git repo (the source of truth), builds to a preview, and goes live only when you say so.

The site's content is plain markdown plus a one-line %% sitetile: … %% grammar in a git repo — an LLM's native tongue, not a wall of block JSON. So an agent connected here can read what it built, not just write blindly. That's the point.

npx @feelreef/mcp        # or add it to your MCP client (below)

Scope, honestly. These tools edit a feelreef native site that already exists — there is no create-a-site-from-nothing tool, and the hosted path for owners who aren't us is still parked (src/phase2/). Building needs either the sitetile renderer on your machine or a runner we provision per site. If you're a cold-start agent hoping to spin a site up unattended: not today. Editing, reading back, previewing and publishing an existing site — that part is real and it's yours. See https://feelreef.com/for/agents.

Install (Claude Code)

claude mcp add feelreef -- npx -y @feelreef/mcp

Point any MCP client at the feelreef-mcp binary. Editing needs nothing but Node ≥20 — the sitetile parser is vendored into the package, so get_page / save_page / create_page work on a fresh machine with no other checkout. Building (turning the repo into a deployed site) needs either the local sitetile renderer or a runner — see Building.

The tools

| tool | what it does | |---|---| | list_sites | the sites this adapter can edit | | add_site / check_site / remove_site | register / preflight / unregister a site | | list_pages · get_page | a site's pages + posts; one page's markdown + parsed structure | | describe_page_grammar | the coral vocabulary for a page body, so you never guess a section | | save_page · create_page · create_post | validate, write, git-commit (no deploy) | | inspect_page | see the built page: links, images, headings + a problems list | | site_status | repo vs live — did my edit actually ship? | | build_preview | build the repo IR → deploy to the preview *.pages.dev | | publish_site | deploy to the live domain. Never automatic; needs confirm:true |

No delete tool by design: an agent that can't delete a page can't lose one.

The loop

get_page → (edit the markdown) → save_page → build_preview → inspect_page the *.pages.dev
                                                                   │
                                                     happy?  →  publish_site {confirm:true}

save_page commits to the repo but changes nothing live. build_preview shows the result on a preview URL. publish_site is the only thing that touches the real domain.

Building

build_preview / publish_site are the same verbs wherever the build runs. The adapter picks the path automatically:

• Local — on a machine with the sitetile renderer (~/Developer/tile), the build runs here. This is the maintainer's dogfood path.

• Runner — on any other machine, point the site at a reef-runner and the build happens in the cloud: the adapter uploads the site's IR, the runner builds and deploys it, and hands back the same honest report (a routes manifest and a read-back that the deployed URL really serves this build). Configure it per site:

  add_site key:"mysite" repo:"/path/to/site" previewProject:"mysite-preview"
           runner:{ url:"https://<your-reef-runner>", token:"<token>" }

• Neither — build_preview returns an honest error telling you to set up one of the above, rather than failing cryptically.

Runner access is provisioned per site today, not self-serve — ask us ([email protected]) and we'll point one at your site. The runner's source and protocol spec live in the monorepo, which isn't public yet; what ships here is the adapter that talks to it.

Adding a site

add_site never trusts a path — it preflights the repo first (is it git, does its IR parse as sitetile, is there a build-fixture trap, does Cloudflare actually bind the domain) and refuses anything unsafe to edit. A site with a live domain + prod project is added full (can publish); a site with only a preview project is preview-only (edit + preview, publish disabled until cutover).

License

MIT © CVER Inc. — including the vendored sitetile parser in vendor/, which ships in this tarball and is the thing that reads and writes your markdown. Read it: that's the whole content model. (The renderer's own repo isn't public yet; the moat was never the parser — it's the loop quality and the running platform.)