MCP overview

The buzzabout MCP server lets any Model Context Protocol client — Claude Desktop, Claude.ai, Claude Code, Codex, Cursor, ChatGPT, your own SDK agent — drive buzzabout from inside the conversation. You ask a research question; the assistant collects and analyses the posts and returns the answer in the chat, often as an interactive widget.

Transport

https://mcp.buzzabout-staging.com/mcp/

Streamable HTTP — one URL handles tool listing, tool calls, and OAuth discovery (no separate SSE endpoint).

See Use in your agent for per-host wiring.

Authentication

Two paths on the same URL, chosen by client:

• OAuth — standard assistants (Claude Desktop, Claude.ai, ChatGPT) handle it; just sign in.
• x-api-key — CLI / IDE agents (Claude Code, Codex, Cursor) and custom agents paste a key.

Both resolve to the same buzzabout account. Full details on Authentication.

Data collection and pattern detection incur plan-priced credit charges. Assistant and natural-language query turns consume variable token inference, with requested data work charged separately. Raw lookups, status reads, and rendering existing results are free. See Pricing.

What's exposed

15 tools, in three kinds:

• Research tools you drive — collect_mentions, get_run, detect_patterns, query. You write the search query, decide when to cluster, and ask your own questions of the result. The reasoning stays with your agent; these tools collect the conversation and hand it back.
• The assistant — ask → get_message → render. Hands the whole job to the buzzabout assistant instead. Use it for what the research tools do not cover: a written report or asset, setting up a tracking agent, saving an insight. Asynchronous — poll get_message, and render draws any rich block as an interactive widget.
• Lookups — collections, mentions, tracking agents and the account.

The async flow

Research can run for minutes — past a host's tool-call timeout — so the assistant flow is asynchronous:

1. buzzabout__ask(prompt) returns immediately with { chat_id, message_id, status: "working" }.
2. The host polls buzzabout__get_message(chat_id, message_id) — which long-polls (one call holds for ~45s, returning the moment the turn settles) — until stop_reason is non-null.
3. The answer comes back as blocks: a render: true block is shown via buzzabout__render as an interactive MCP App widget; a render: false block carries plain markdown the host relays. Hosts without the MCP Apps UI extension get markdown throughout.

An exhausted balance prevents new paid work. Handle the returned structured error rather than guessing its units from the machine code. Credit presentation uses explicitly named credit projections; existing USD fields remain USD. Top up in the app before retrying. Calling ask again is a new charged turn, not a free status check.

When to use MCP vs REST

Both surfaces are backed by the same primitives, so hybrid setups work naturally (e.g. run a heavy collection via REST, then ask an MCP-capable assistant to summarise it).

See also

• Agentation — third-party MCP devtool that pairs well with buzzabout for local agent development and inspection.

Next steps

• Tools reference — the 15-tool surface.
• Use in your agent — per-host wiring + first prompt.
• Authentication — pick OAuth or x-api-key.