Squad MCP Server

A remote MCP server that brings Squad — the AI product feedback intelligence platform — directly into your AI workflows. Connect Squad to Claude, ChatGPT, or any MCP-compatible AI assistant to turn raw user feedback into signals, insights, actions, and briefs without context switching.

Squad continuously ingests feedback, clusters it into signals, distils it into insights, and links it to the actions and goals that move your product forward. The MCP server exposes that same intelligence — read the evidence behind a decision, capture new feedback, and generate briefs from your assistant.

🚀 Quick Start

For Users

Connect Squad to your AI assistant in seconds:

Claude Code:

claude mcp add --transport http squad https://mcp.meetsquad.ai/mcp

On first use, you'll be prompted to authenticate via OAuth in your browser.

Other MCP Clients:

Connect using `https://mcp.meetsquad.ai/mcp` — OAuth configuration is automatically discovered via the server's `.well-known/oauth-protected-resource` metadata (which points clients at PropelAuth as the authorization server).

📖 Usage Examples

See **USAGE_EXAMPLES.md** for detailed real-world examples. A few things you can ask:

• Triage feedback — "Capture this support ticket in Squad and tell me if it's a known theme."
• Weekly review — "Run my weekly product review: what changed and what needs deciding?"
• Ground the evidence — "Show me the customer signals behind insight IN-42."
• Draft a brief — "Generate a brief for action AC-12."
• Search everything — "Find all feedback related to onboarding friction."
• Ground a ticket — "Pull the customer evidence behind AC-7 before I build it."

Squad entities are referenced by short display IDs so the assistant can cite its evidence:

Tools

The server exposes 30 tools. Write tools require a token minted with the `write:workspace` scope; read tools only need `read:workspace`.

• `list_workspaces` — List every organisation and workspace you can access, with the current selection marked.
• `select_workspace` — Select which organisation and workspace subsequent tools operate on.
• `get_workspace_overview` — One-call orientation: mission and description, top goals, recent signal activity, evidence-chain health, and open work counts.
• `update_workspace` — Update the current workspace's name, description, mission statement or logo.
• `list_members` — People in the current organisation with their user IDs.
• `search` — Keyword search across signals, insights, actions, goals, documents and clusters.
• `get_entity` — Fetch any entity by display ID (`SI-1`, `IN-1`, `AC-1`, `GL-1`, `BR-1`, `DC-1`, `CL-1`) or UUID. Also how you check on async work.
• `list_signals` — Browse raw feedback signals with filters for source, type, sentiment, cluster and date range.
• `find_similar_signals` — Semantically related signals for a given signal — "has anyone else said this?".
• `list_clusters` — Browse signal clusters (recurring themes in feedback) with sizes and labels.
• `get_cluster` — A cluster's label, stats, member signals and linked insights.
• `list_insights` — Browse distilled insights ranked by combined score, with category/score/status filters or scoped to a goal.
• `list_actions` — Ranked actions (what the evidence says to do next) filtered by status, assignee, priority or parent insight.
• `get_action_context` — Everything needed to execute an action in one call: the action, the parent insight, the customer evidence behind it, and the goals it serves.
• `update_action` — Edit an action's priority, effort, category or notes; assign it; or link it to an insight or brief.
• `update_action_status` — Move an action through its lifecycle: start, complete, dismiss or snooze.
• `list_goals` — Strategic goals ordered by importance.
• `create_goal` — Create a strategic goal.
• `update_goal` — Update a goal's title, description or importance.
• `update_insight` — Curate an insight: set category or status, and link/unlink the goal it supports.
• `dismiss_signal` — Permanently remove a signal from the workspace (noise, spam, or mis-ingested content).
• `get_activity` — The workspace change feed (humans and Squad agents), newest first.
• `list_documents` — Browse workspace knowledge documents (and briefs) with their paths and tags.
• `create_document` — Create a knowledge document from markdown (research summaries, meeting notes, analyses).
• `update_document` — Replace a document's markdown body and/or title, and manage tags.
• `list_briefs` — Briefs with their status (building/draft/in_review/finalised/failed) and recommendation.
• `generate_brief` — Kick off AI generation of a brief from an action or insight.
• `update_brief_status` — Move a brief through its review lifecycle: draft, in_review, or finalised.
• `ingest_signal` — Pipe user feedback into the evidence chain (1–50 items, deduplicated server-side).
• `list_integrations` — Connected feedback sources for this workspace and their sync health.

Prompts

Ready-made workflows exposed as MCP prompts:

• `triage-feedback` — check for duplicates, ingest a piece of feedback, and report where it landed.
• `weekly-product-review` — what changed, what the evidence says, and what needs deciding.
• `draft-decision-brief` — generate a brief from an action or insight and walk it to a readable draft.
• `ground-this-ticket` — for coding agents: pull the customer evidence behind a piece of work before building it.

Resources

Pin these so strategy questions need no tool calls:

• `squad://workspace/context` — the current workspace's mission and product context.
• `squad://goals` — the workspace's strategic goals with importance rankings.

Tool Capabilities

• ✅ Safety annotations (`readOnlyHint` / `destructiveHint`) on every tool
• ✅ Structured Zod input schemas
• ✅ User- and workspace-isolated data access via OAuth
• ✅ Scope-gated writes (`write:workspace`)

🏗️ Architecture

┌─────────────┐         OAuth          ┌──────────────┐
│  Claude /   │ ◄────────────────────► │  PropelAuth  │
│  ChatGPT    │    (Authentication)     │   (IdP)      │
└─────────────┘                         └──────────────┘
       │
       │ HTTPS + Bearer Token
       ▼
┌──────────────────────────────────────────────┐
│  Squad MCP Server                            │
│  ┌────────────────────────────────────────┐  │
│  │  OAuth → introspect + verify token     │  │
│  │  JWT minting → service credentials     │  │
│  │  Redis → workspace selection + tokens  │  │
│  │  MCP handler → tools / prompts / res.  │  │
│  │  PostHog → tool-call telemetry         │  │
│  └────────────────────────────────────────┘  │
└──────────────────────────────────────────────┘
       │
       │ Squad API Calls (minted JWT)
       ▼
┌──────────────┐
│  Squad API   │
└──────────────┘

The server is built on `mcp-use` v2 and talks to the Squad platform API over GraphQL. Each request is served statelessly: the MCP layer holds no session, and every call re-introspects its own bearer token. Redis stores the durable per-user state (workspace selection and minted-token cache), so any instance can serve any request. Backend types are generated from a committed GraphQL schema snapshot (see GraphQL codegen).

🛠️ Development

This repository contains the source code for the Squad MCP remote server.

Prerequisites

• Node.js 22+
• pnpm
• Nix (optional, for a reproducible dev environment via `flake.nix`)
• PropelAuth credentials (OAuth 2.1 client + backend API key)
• Redis (optional locally; workspace selection falls back to in-memory)

Local Setup

# Clone repository
git clone https://github.com/the-basilisk-ai/squad-mcp.git
cd squad-mcp

# Install dependencies
pnpm install

# Configure environment
cp .env.example .env
# Edit .env with your PropelAuth credentials (and SQUAD_ENV=dev to target the dev platform)

# Start development server with hot reload
pnpm dev

# Server available at http://localhost:3232

Environment Variables

Available Commands

pnpm dev                # Start dev server with hot reload (mcp-use)
pnpm build              # Build the server (mcp-use)
pnpm start              # Start the built server
pnpm deploy             # Deploy via mcp-use
pnpm test               # Run unit tests (vitest)
pnpm format             # Lint/format check (biome)
pnpm format:fix         # Auto-fix lint/format issues
pnpm codegen            # Regenerate GraphQL types from schema.graphql
pnpm codegen:check      # Fail if generated GraphQL types are stale

Testing the Server

# Check health
curl http://localhost:3232/health

# Check OAuth discovery
curl http://localhost:3232/.well-known/oauth-protected-resource

# Test with the built-in inspector
pnpm dev   # then open the inspector and connect to http://localhost:3232/mcp

Project Structure

squad-mcp/
├── server.ts                   # MCP server entry point (OAuth, tool/prompt/resource registration)
├── server.json                 # MCP registry metadata (see MCP_REGISTRY.md)
├── schema.graphql              # Committed snapshot of the Squad platform GraphQL schema
├── codegen.ts                  # GraphQL Code Generator config
├── src/
│   ├── tools/                  # Tool implementations, grouped by surface
│   │   ├── registry.ts         # Single registration path (annotations, errors, telemetry)
│   │   ├── workspace.ts        # list/select workspaces, overview, members
│   │   ├── search.ts           # semantic search
│   │   ├── get-entity.ts       # fetch any entity by display ID / UUID
│   │   ├── evidence.ts         # signals, clusters, insights
│   │   ├── actions-read.ts     # list actions, action context
│   │   ├── actions-write.ts    # update actions + status
│   │   ├── strategy-read.ts    # goals, activity
│   │   ├── strategy-write.ts   # create/update goals, insights, dismiss signals
│   │   ├── knowledge.ts        # documents + briefs
│   │   ├── ingest.ts           # ingest new signals
│   │   └── integrations.ts     # list connected sources
│   ├── prompts/                # MCP prompt workflows
│   ├── resources/              # MCP resources (workspace context, goals)
│   ├── gql/                    # Generated GraphQL types (pnpm codegen)
│   ├── graphql/                # GraphQL operation documents
│   ├── helpers/                # OAuth provider, token minting, workspace selection, KV/Redis
│   └── lib/                    # Squad API client, logger, telemetry
├── railway.toml                # Railway deployment config
└── .env.example                # Environment template

🏭 Production Deployment

This is a hosted service maintained by Squad. Users connect via OAuth — no self-hosting required.

Architecture notes (for contributors):

• Deployed on Railway with a `/health` readiness check
• Stateless request handling, with Redis holding the per-user workspace selection and token cache, so instances scale horizontally
• Follows the MCP specification for streamable HTTP transport

💬 Support

Need help with the Squad MCP server?

• Email: support@meetsquad.ai
• Documentation:
  • Squad MCP Guide — complete setup and integration guide
  • USAGE_EXAMPLES.md — real-world usage examples
• Issues: GitHub Issues — bug reports and feature requests
• Privacy Policy: meetsquad.ai/privacy-policy
• Squad Platform: meetsquad.ai

🤝 Contributing

Contributions welcome! Pre-commit hooks run biome and vitest automatically. Please ensure:

• `pnpm format` passes (biome)
• `pnpm build` compiles without errors
• `pnpm test` passes
• `pnpm codegen:check` passes if you touched GraphQL operations
• All tools include safety annotations

📄 License

MIT

🔗 Links

• Squad MCP Documentation — complete setup and integration guide
• Squad Platform
• MCP Specification
• Issue Tracker

GraphQL codegen

Backend access is typed via GraphQL Code Generator. `schema.graphql` is a

committed snapshot of the Squad platform API schema; `src/gql/` is generated

from it plus the operation documents in `src/graphql/`.

• Refresh the snapshot: copy `packages/graphql/src/schema/generated.graphql`

from the API repo over `schema.graphql` (or set `SQUAD_GRAPHQL_URL` to

introspect a live endpoint), then run `pnpm codegen`.

• CI runs `pnpm codegen:check` and fails when `src/gql/` is stale.