AgentDrop

Give your agents a simple way to host static sites.

AgentDrop turns a single HTTP call into a live, shareable URL — Markdown or HTML — that is versioned, undoable, real-time, and requires no account. It's built for AI coding agents: install one skill, and your agent can publish and update sites on its own.

Built on a Turborepo monorepo with TanStack Start, Convex, and COSS UI. Site content lives in Cloudflare R2 (via the @convex-dev/r2 component), version history / undo is handled by convex-timeline, and abuse protection by @convex-dev/rate-limiter.

Quickstart (for your agent)

npx skills add opencoredev/agent-drop

That installs skills/agentdrop/SKILL.md into whichever agents you pick. Regenerate that file after changing packages/backend/convex/skill.ts, or when pointing at a new deployment:

bun run skill:sync                                        # uses CONVEX_SITE_URL
AGENTDROP_API_BASE=https://prod.convex.site bun run skill:sync

Or call the API directly:

curl -X POST https://<your-convex-site>/api/v1/sites \
  -H "Content-Type: application/json" \
  -d '{"kind":"markdown","title":"Hello","content":"# Hi from my agent"}'
# → { "url": "...", "manageUrl": "...", "editToken": "..." }

Save the editToken to update/undo/delete later. Point your agent at the hosted skill (<your-convex-site>/agentdrop-skill.md) and it handles the rest.

How it works

agent ── POST /api/v1/sites ──▶ Convex httpAction
                                  ├─ store content blob in R2  (@convex-dev/r2)
                                  ├─ push a version snapshot   (convex-timeline)
                                  └─ upsert `sites` row (currentKey, editTokenHash, expiresAt)
viewer ── useQuery(getBySlug) ──▶ reactive metadata + content URL
                                  └─ re-renders instantly on any deploy / update / undo

• No site storage in Convex. Convex holds only metadata + R2 keys; blobs + images live in R2.
• No account by default. Deploy returns a secret editToken (stored only as a SHA-256 hash) that authorizes updates/undo/delete.
• Retention. Anonymous sites: 30 days. Signed-in (claimed) sites: 90 days. Images: 7 days. A daily cron purges expired R2 objects + rows + timeline scopes.
• Security. Content is scanned for obvious credentials; untrusted HTML is served from the Convex .site origin with a sandbox CSP and rendered inside a sandbox="allow-scripts" iframe.

The API

Method & path	Auth	Purpose
POST /api/v1/sites	—	Deploy a site → { slug, url, manageUrl, editToken }
PUT /api/v1/sites/:slug	Bearer	Replace content (keeps history)
POST /api/v1/sites/:slug/undo · /redo	Bearer	Step through history
GET /api/v1/sites/:slug	—	Status (canUndo/canRedo/version)
GET /api/v1/sites/:slug/raw	—	Raw content
POST /api/v1/sites/:slug/images	Bearer	Upload an image (≤5 MB, ≤10/site)
DELETE /api/v1/sites/:slug	Bearer	Delete the site
GET /agentdrop-skill.md	—	The distributable agent skill
POST /mcp	—	Stateless MCP server (Streamable HTTP)

The agent-facing reference is served live at <convex-site>/agentdrop-skill.md.

MCP

POST /mcp speaks MCP over Streamable HTTP, revision 2026-07-28: no protocol-level session, no GET/SSE stream, and no initialize handshake. Every POST carries its own protocol version, client info, and capabilities in params._meta, mirrored into the MCP-Protocol-Version, Mcp-Method, and Mcp-Name headers, which the server validates against the body. Holding no state between calls is what lets it run on a serverless function with no sticky routing. Clients on 2025-03-26 through 2025-11-25 still get the legacy initialize flow, and no session id is ever minted.

Tools: deploy_site, update_site, undo_site, redo_site, get_site, delete_site.

Connecting a harness (the landing page renders these with the live URL filled in):

claude mcp add --transport http agentdrop <site>/mcp
codex mcp add agentdrop --url <site>/mcp
gemini mcp add --transport http agentdrop <site>/mcp
code --add-mcp '{"name":"agentdrop","type":"http","url":"<site>/mcp"}'

Cursor (~/.cursor/mcp.json) has no CLI for this, so it takes JSON:

{ "mcpServers": { "agentdrop": { "url": "https://<your-convex-site>/mcp" } } }

Project structure

agent-drop/
├── apps/
│   └── web/         # TanStack Start front end (landing, viewer, manage, auth)
├── packages/
│   ├── ui/          # COSS UI primitives (Base UI) + shared styles
│   └── backend/     # Convex: schema, sites functions, HTTP API, R2/timeline/rate-limiter, crons

Setup

bun install

1. Convex

bun run dev:setup   # log in, create the "agent-drop" project, codegen

2. Cloudflare R2

wrangler r2 bucket create agent-drop

Create an R2 API token (dashboard → R2 → Manage R2 API Tokens → Object Read & Write), then set the Convex environment variables:

cd packages/backend
bunx convex env set R2_BUCKET agent-drop
bunx convex env set R2_ENDPOINT https://<ACCOUNT_ID>.r2.cloudflarestorage.com
bunx convex env set R2_ACCESS_KEY_ID <ACCESS_KEY_ID>
bunx convex env set R2_SECRET_ACCESS_KEY <SECRET_ACCESS_KEY>
bunx convex env set R2_TOKEN <TOKEN_VALUE>
bunx convex env set SITE_URL http://localhost:3001
bunx convex env set BETTER_AUTH_SECRET "$(openssl rand -base64 32)"

Copy VITE_CONVEX_URL / VITE_CONVEX_SITE_URL from packages/backend/.env.local into apps/web/.env.

3. Run

bun run dev          # web on http://localhost:3001 + convex dev

Deploying

bun run deploy          # Convex first, then the web app

Deploy the backend before the frontend, and use this script rather than deploying from the Vercel dashboard alone. apps/web/vercel.json sets its own buildCommand, which overrides the project's dashboard build command, so a web deploy does not push Convex; deploying only the web app leaves the API on the previous release. (vercel.json rejects unknown keys, including comment keys, so that warning lives here rather than in the file.)

Scripts

• bun run dev — all apps in dev
• bun run dev:web / bun run dev:server — web / Convex only
• bun run check-types — typecheck all packages
• bun run check — Oxlint + Oxfmt