Prefixed identifiers — `SI-` for a signal, `IN-` for an insight, `AC-` for an action — let a model cite the exact evidence behind a claim instead of summarising it into something unverifiable. That is the difference between a product assistant you can check and one you have to trust. The `ground-this-ticket` prompt points at the same idea from the other end: before a coding agent builds something, pull the customer evidence for why. Writes are scope-gated, so a read-only token is a real option.
A hosted MCP server for Squad, which ingests user feedback, clusters it into signals, distils insights, and links those to the actions and goals a product team works on. The server exposes around 35 tools over that model, and every entity has a short display id so an assistant can cite what it is claiming.
- Orientation — `list_workspaces`, `select_workspace`, `get_workspace_overview`, `update_workspace`, `list_members`
- Retrieval across everything: `search` for semantic search and `get_entity` to fetch any object by its display id or UUID
- The evidence layer — `list_signals`, `find_similar_signals`, `list_clusters`, `get_cluster`, `list_insights`
- Work tracking through `list_actions`, `get_action_context`, `update_action` and `update_action_status`
- Strategy — `list_goals`, `create_goal`, `update_goal`, `update_insight`, `dismiss_signal`, `get_activity`
- `ingest_signal` — captures new feedback with deduplication, so a support ticket lands as a signal and you are told whether it is a known theme
- Briefs and knowledge: `generate_brief`, `list_briefs`, `update_brief_status`, plus `list_documents`, `create_document`, `update_document`
- Four prompts that compose the tools — `triage-feedback`, `weekly-product-review`, `draft-decision-brief`, and `ground-this-ticket` for coding agents
- Two pinnable resources: `squad://workspace/context` for mission and product context, `squad://goals` for ranked strategic goals
- Safety annotations on every tool, marking which are read-only and which are destructive
A Squad account. Connect with `claude mcp add --transport http squad https://mcp.meetsquad.ai/mcp` and complete OAuth in the browser on first use; other clients discover the OAuth configuration automatically from the server's protected-resource metadata. Nothing runs locally. Scopes matter: read tools need `read:workspace` and write tools need a token minted with `write:workspace`. Self-hosting is not expected — this is a hosted service — but the source is here, built on Node.js 22+ with pnpm. MIT licensed.
One command — claude mcp add --transport http squad https://mcp.meetsquad.ai/mcp