Interfaces

MCP Server

Streamable-HTTP MCP server for Claude Code integration.

The Drop MCP server lets Claude Code (and any MCP-compatible client) manage Droplets directly — creating, searching, editing, sharing, and versioning without leaving the conversation. No CLI install required on the agent's machine.

Drop ships two transports:

Remote (recommended) — Streamable HTTP at https://drop.himarley.ai/api/mcp. Bearer-token auth, stateless, no install. Use this in Claude Code via "type": "http".
Local stdio — npx @himarley/drop-mcp. Adds support for reading local file paths (filePath arg) which the remote server can't do — useful when authoring a Droplet from a file on disk.

The CLI is still the most token-efficient surface for Claude Code; the MCP server is the right pick for non-CLI MCP clients or when you want tool-discovery without a shell.

Available tools

Tool	What it does
`drop_list`	List your Droplets (scope mine/shared/all, sort, order, limit)
`drop_get`	Get details for a specific Droplet
`drop_search`	Full-text search across all Droplets
`drop_create`	Create a Droplet from HTML or Markdown
`drop_redeploy`	Update an existing Droplet with new content
`drop_download`	Download the HTML source of any version
`drop_versions`	View version history
`drop_rollback`	Roll back to a previous version
`drop_duplicate`	Copy a Droplet to your own account
`drop_get_access`	View internal access settings
`drop_set_access`	Set PRIVATE / EVERYONE / SPECIFIC / SPACE access
`drop_share_links`	List external share links
`drop_create_share_link`	Create an internal or external share link
`drop_delete_share_link`	Remove a share link
`drop_comments`	List threaded comments on a Droplet (legacy shape)
`drop_comments_create`	Post a comment — page-level, pinned to an element, or a threaded reply
`drop_comments_list`	List comments with pin target, resolved state, and thumbs-up counts
`drop_comments_get`	Fetch a single comment by id
`drop_comments_resolve` / `drop_comments_unresolve`	Close or reopen a comment — resolve is the AI-loop close verb, not delete
`drop_comments_delete`	Permanently remove a comment + its replies (spam / mistakes only)
`drop_comments_thumbs_up` / `drop_comments_unthumbs_up`	Add or remove your thumbs-up reaction
`drop_state_get`	Read a Droplet's shared key/value state
`drop_state_set`	Write a Droplet's shared state — modify it directly, no JSON import step
`drop_signoff_list`	List decisions you can see, with sign-off state; filter by phase/role/tag
`drop_signoff_get`	Read a decision's sign-off — state + the actions available to you
`drop_signoff_cast`	Cast your sign-off (agree / condition / disagree)
`drop_signoff_clear`	Remove your own sign-off response while the decision is open
`drop_signoff_finalize`	Record the decider's call (approve → decided, block → blocked)
`drop_signoff_reopen`	Clear a recorded call and reopen
`drop_signoff_open` / `drop_signoff_close`	Open a draft decision for sign-off, or return it to draft (editor)
`drop_signoff_set_proposition`	Set "what you're signing off on" (editor)
`drop_signoff_mine`	List the decisions across Drop that are waiting on you

Operations

Configure Claude Code

Generate a token at Settings → API Tokens (it's only shown once — copy it). Then add the MCP server to Claude Code via /settings. Restart Claude Code and it auto-discovers every tool.

{
  "mcpServers": {
    "drop": {
      "type": "http",
      "url": "https://drop.himarley.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer drop_your_token_here"
      }
    }
  }
}

Authenticate

The remote MCP server accepts the same Bearer tokens the CLI uses — either a Cognito JWT (issued by drop login) or a legacy drop_* token (deprecated but still accepted during the migration window). API tokens are managed at /settings/tokens. A PR preview (engineering) needs a token minted inside that preview: eval "$(drop preview token <PR>)" sets DROP_SERVER and DROP_TOKEN for it, and drop preview exec <PR> -- <command> runs a local MCP server (the PR's own, say) with them and an empty HOME. The local MCP server refuses a drop login token aimed at a PR preview, and refuses DROP_SERVER without DROP_TOKEN when it names a server other than your saved login's.

# Confirm a token works against the MCP endpoint
curl -X POST https://drop.himarley.ai/api/mcp \
  -H "Authorization: Bearer drop_your_token_here" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Call a tool

Each tool follows the standard MCP tools/call shape. Sample prompts that land cleanly: "Build a team dashboard and publish it as team-dashboard", "Roll my-dashboard back to v3", "Search Drop for anything related to onboarding", "Share the quarterly-report droplet with all of Hi Marley".

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "drop_search",
    "arguments": { "query": "onboarding" }
  }
}