Summarize a busy channel

The intelligence_summarize_channel tool turns “what did I miss in #support?” into a structured summary - without the server holding any LLM API key. It does this through MCP sampling: the tool calls back to the host client’s LLM via the requestSampling capability, then parses the response.

Use case

A support engineer says: “TL;DR what happened in #support over the last 50 messages, then post the summary to #ops.”

Two tool calls. The first asks the host’s own LLM (Claude, GPT, whatever) to do the heavy lifting; the second posts the result. discord-mcp ships zero API keys and never makes outbound LLM requests on your behalf.

Note

Sampling capability is negotiated at initialization. Do not infer support from a client name or version. When sampling is absent, the tool returns raw messages plus _meta.fallback: 'host_llm_should_process'; the host model must produce the summary in a later turn. The fallback result is data to summarize, not a completed summary.

Tool flow

1. Read the channel (optional).

   intelligence_summarize_channel fetches messages internally, so this step is only needed if you want to inspect the raw stream first or pass a custom subset. For the standard flow, skip straight to step 2.

   {
     "name": "messages_read",
     "arguments": {
       "channel_id": "222233334444555566",
       "limit": 50
     }
   }

2. Summarize.

   The tool returns a structured object with summary, key_topics, and action_items. The style parameter controls tone: bullet (default), paragraph, or executive.

   Input

   {
     "name": "intelligence_summarize_channel",
     "arguments": {
       "channel_id": "222233334444555566",
       "limit": 50,
       "style": "bullet"
     }
   }

   Output (sampling supported)

   {
     "summary": "- Three users hit the new SSO flow with expired refresh tokens.\n- @alice shipped a hotfix at 14:22 UTC; @bob verified.\n- Open: customer #88241 still cannot log in - escalated to on-call.",
     "key_topics": ["sso", "refresh tokens", "hotfix deploy"],
     "action_items": ["follow up with customer #88241"],
     "message_count_used": 50,
     "sampling_used": true
   }

   Output (fallback)

   {
     "message_count_used": 50,
     "channel_id": "222233334444555566",
     "style": "bullet",
     "raw_messages": [
       { "id": "888899990000111101", "author": "alice", "content": "shipping the SSO hotfix now" },
       { "id": "888899990000111102", "author": "bob", "content": "lgtm, will verify on staging" }
     ],
     "_meta": {
       "fallback": "host_llm_should_process",
       "intent": "summarize",
       "sampling_used": false
     }
   }

3. Post the summary.

   {
     "name": "messages_send",
     "arguments": {
       "channel_id": "333344445555666677",
       "content": "## Support recap\n\n- Three users hit the new SSO flow with expired refresh tokens.\n- @alice shipped a hotfix at 14:22 UTC; @bob verified.\n- Open: customer #88241 still cannot log in - escalated to on-call."
     }
   }

   When the tool returned the fallback envelope, the host LLM should instead read the top-level raw_messages, generate the summary itself in the next turn, then call messages_send with that content.

Why this design

• Zero API keys to manage. The server has no OPENAI_API_KEY field, no budget telemetry, no key rotation playbook. The agent’s existing model subscription does the work.
• Privacy. Channel content never leaves the path the user already trusts (their MCP client → their LLM). discord-mcp doesn’t proxy text to a third-party inference provider.
• Explicit degradation. Clients without sampling receive the source data and a machine-readable fallback marker. The host can then summarize it in a separate model turn.

Style options

style	Output shape
bullet (default)	Markdown bullet list, scannable
paragraph	Two-to-three sentence paragraph
executive	One business-focused paragraph

See also

• intelligence_summarize_channel
  • full schema and fallback envelope.
• messages_read - inspect raw messages before summarizing.
• messages_send - post the summary back.
• intelligence_classify_messages, intelligence_moderate_content
  • sibling sampling-based intelligence tools.
• Architecture: MCP sampling - how the fallback envelope is constructed.