Skip to content
SparkleTree SparkleTree Docs

SparkleTree for agents

You are reading documentation written for AI agents. SparkleTree is a context-aware marketing platform: human marketers (and you, on their behalf) create campaigns whose core creative unit is the Squid, a living, conversational ad that greets, talks, and adapts. Campaigns hold variants; an adaptive engine (Adaptive CVG) learns which variant performs and shows winners more often. Campaigns deploy to surfaces (websites, landing pages, physical screens, kiosks, QR codes, Meta ads), and every published brand gets a public brand page at its own domain or <name>.sparkletree.io. Analytics distinguish Impressions (raw delivery count) from Reach (distinct sessions).

Two integration surfaces

Section titled “Two integration surfaces”

A. Platform MCP server: operate a workspace (auth required). Endpoint: https://app.sparkletree.io/api/mcp. Streamable HTTP, OAuth 2.1 + PKCE with dynamic client registration; a bare URL is all a spec-compliant MCP client needs. Requires an existing SparkleTree account. There is no anonymous or agent-driven signup; a human signs up first. Every tool is scoped to the authenticated user’s workspace and role; cross-workspace access is refused. The default profile exposes 15 consolidated tools: eight read-only query_* tools plus get_platform_info, four manage_* write tools, and three tools that spend credits (generate_campaign, render_video, video_commission). Other tool families sit behind a call code (?profile=<call-code>). Read the mosaic://capabilities resource first; it is the live matrix of tools, prompts, and recommended flows. Rate limit: 120 requests per minute per token.

B. Brand-host public surface: consume brand content (no auth). Every published brand site serves machine-readable surfaces: /llms.txt (entry map), /.well-known/agent-card.json, POST /ask (the brand’s conversational concierge), markdown twins of pages via Accept: text/markdown, /.well-known/agent-skills/index.json (three HTTP-fetchable, sha256-digested skills: ask-venue, fetch-campaign-content, discover-campaigns), /.well-known/mcp/server-card.json, /auth.md, sitemap.xml, robots.txt, and JSON-LD. If your task is answering questions about a brand or fetching its published content, use these: no account, no OAuth, no MCP session.

Rule of thumb: operating (create, edit, publish, analyze) is surface A. Reading a brand’s public content is surface B.

Files in this folder

Section titled “Files in this folder”
  • install.md: machine-consumable install guide. Endpoint, auth paths, profiles, the 15-tool table, resources, prompts, brand-host discovery chain.
  • ready-prompt.md: the orientation prompt a human pastes to their agent after connecting. If a user gave you that prompt, follow it.
  • example-prompts.md: realistic user asks mapped to the tools they exercise.
  • tips-and-tricks.md: operational guidance. Capability discovery, pagination, async polling, backoff, when to prefer public surfaces.
  • advanced-workflows.md: multi-tool recipes. URL-to-live-campaign, weekly analytics digest, video from a winner, no-auth brand Q&A, multi-workspace agency ops.