AI Research Agent

August 26, 2026 · View on GitHub

An AI agent for market and audience research — web research with citations, document ingestion for RAG, and audience/persona profiling — backed by real web-search and document APIs.

Part of Agency Agents OS, an open ecosystem of specialized AI agents for real business work.

What this covers

This repo is the umbrella for anything an agency or in-house team would call "the AI research agent": answering a research question from live web sources with citations, turning PDFs and long documents into structured Markdown for RAG, and profiling a target audience's demographics, pain points, and language for a market/persona brief.

Sub-agents

AgentDoesStatus
Web ResearchSynthesizes an answer to a research question from live web search across multiple sources, with citationsComing Soon
Document IngestionConverts PDFs and long documents into structured, RAG-ready Markdown with preserved structureComing Soon
Audience ResearchProfiles a target audience's demographics, pain points, and language for a market/persona briefComing Soon

Required Muapi APIs

  • research.web_search — live multi-source web search and synthesis with citations.
  • research.pdf_to_markdown — structured document-to-Markdown conversion for RAG pipelines.
  • research.audience_profile — audience/persona profiling from web and market signals.

See each sub-agent's SKILL.md for the specific capabilities it uses.

Setup

  1. Create a Muapi account and API key at muapi.ai.
  2. Review the Muapi API quickstart and OpenAPI schema for the research endpoints.
  3. Load the SKILL.md for the sub-agent you need into your agent runtime (hosted agent, MCP client, or custom LLM app), or follow it manually.

Using with an AI agent

Every sub-agent's SKILL.md is model- and runtime-agnostic — it's plain Markdown, so it works with any LLM agent, not just Claude. Two integration paths:

As an MCP connection (the agent gets live Muapi tools):

Muapi runs an MCP server at https://api.muapi.ai/mcp that any MCP-compatible client can connect to — Cursor, Windsurf, Claude, or your own custom agent.

  • Cursor / Windsurf / other clients with a header field: connect to https://api.muapi.ai/mcp with an Authorization: Bearer YOUR_MUAPI_KEY header.
  • claude.ai / Claude Cowork / other connector UIs with no header field: use the URL-embedded key form instead, https://api.muapi.ai/mcp/YOUR_MUAPI_KEY, via Settings → Connectors → Add custom connector.
  • Claude Code / Claude Desktop: claude mcp add muapi -e MUAPI_API_KEY=YOUR_MUAPI_KEY -- muapi mcp serve (uses the muapi CLI's stdio transport — Claude Code's HTTP MCP client doesn't reliably inject tools).

Full setup details for every client: muapi.ai/docs/mcp.

As agent instructions (any LLM follows the workflow directly):

Drop a sub-agent's SKILL.md into a Claude Code project's .claude/skills/ directory, paste it into a custom-GPT/Project's system instructions, hand it to an autonomous agent framework as a tool spec, or attach it directly in a chat conversation — then ask the agent to follow it.

Read-only vs. write actions

All actions in this repo are read-only. Research, document conversion, and audience profiling produce files, Markdown, or reports — none of them write to or publish on any external system. Acting on the research (publishing, outreach, ad targeting) is out of scope; hand the output to the relevant agent for that step, with explicit approval.

Status and limitations

All three sub-agents are Coming Soon. They depend on research.web_search, research.pdf_to_markdown, and research.audience_profile, none of which are yet live on Muapi. This repo documents the intended interface so agents and workflows can be built against it ahead of launch.

Contributing

See Agency Agents OS CONTRIBUTING.md.

License

MIT