Install AgentGem's memory-and-coach features into a Buzz community. @mention AgentGem in any channel and it routes your question to the right specialist — answering with receipts drawn from your team's past agent and coding sessions.
buzz install github:ninemindai/agentgem-buzz-pack
# or a pinned version
buzz install github:ninemindai/agentgem-buzz-pack@v0.1.0| Persona | Answers | AgentGem tools |
|---|---|---|
| AgentGem (orchestrator) | routes the question, synthesizes | — |
| Historian | "where / why did X happen" | search_session_content, ask_session |
| Scorekeeper | "how did session Z go" | summarize_session, search_sessions |
| Coach | "how do we improve" | get_behavior_findings |
| ReportCard | "score my steering / give me a report card" | gemit skill (runs the CLI) |
The tools come from the agentgem MCP server (the @ninemind/agentgem npm package, agentgem-goldmine bin), spawned on the agent host via npx. It reads the local AgentGem recall index at ~/.agentgem/recall-index.db.
So this pack is meaningful where the Buzz agent runs on a machine that has AgentGem installed and indexed (a developer's box, or a self-hosted relay host). Community-wide recall — searching the whole channel's history rather than one host's index — is a planned follow-up that points the recall index at Buzz relay events instead of local transcripts.
Historian, Scorekeeper, and Coach are read-only. Their six goldmine tools read an already-redacted local index; none write. ask_session reads a transcript with a separate agent and returns only the answer — the raw transcript never enters the channel.
ReportCard can execute a command — scoped to that one persona. The gemit report card is a CLI (npx … gemit), not a read-only tool, so ReportCard is granted Buzz's shell MCP (buzz-dev-mcp) via its own mcp_servers front matter. That is a broad capability (a shell can run anything), deliberately isolated to ReportCard and absent from the other three personas. Only add this pack to a community where you're comfortable with that persona running commands on the agent host.
- ReportCard runs
gemitfor the score and only runsgemit --share(which publishes an unlisted, redacted card) when someone explicitly asks — it never passes--yeson your behalf. - Cleaner future path: expose
gemitas a proper read-only-style MCP tool in theagentgem-goldmineserver (arun_gemittool). That would drop the shell requirement entirely — tracked as a follow-up in the AgentGem repo.
MIT © NineMind