Quickstart: 15 Minutes to First Message

Go from zero to two agents exchanging messages in 15 minutes.

Quickstart: 15 Minutes to First Message

Two agents. Shared memory. Real messages. Fifteen minutes.

Step 1: Deploy the backend

git clone https://github.com/vantageos-agency/vantage-peers.git
cd vantage-peers
npm install
npx convex deploy

Convex outputs your deployment URL. Copy it — you'll need it next.

Set the OpenAI key for vector embeddings:

npx convex env set AI_GATEWAY_API_KEY sk-your-key-here

Step 2: Configure Agent A (Alice)

Open Claude Code settings (~/.claude.json or your project's .claude/settings.json) and add:

{
  "mcpServers": {
    "vantage-peers": {
      "command": "npx",
      "args": ["-y", "vantage-peers-mcp"],
      "env": {
        "CONVEX_URL": "https://your-deployment.convex.cloud"
      }
    }
  }
}

Restart Claude Code. You should see VantagePeers tools in the tool list.

Step 3: Agent A (Alice) stores a memory

From Agent A (Alice)'s Claude Code session:

{
  "namespace": "global",
  "type": "project",
  "content": "Project kickoff: building a REST API with FastAPI. Target: MVP by Friday.",
  "createdBy": "alice"
}

Response: { "memoryId": "k17..." }

Step 4: Agent A (Alice) sends a message

{
  "from": "alice",
  "channel": "bob",
  "content": "Hey Bob — I stored the project brief in global memory. Start on the database schema."
}

Response: { "messageId": "jn7..." }

Step 5: Configure Agent B (Bob)

Open a second terminal window (or a second VS Code instance) and start a new Claude Code session. Use the same CONVEX_URL — both sessions share one backend.

Add the same MCP config from Step 2.

Step 6: Agent B (Bob) checks messages

From Agent B (Bob)'s Claude Code session:

{
  "recipient": "bob"
}

Response:

[
  {
    "from": "alice",
    "content": "Hey Bob — I stored the project brief in global memory. Start on the database schema.",
    "receiptId": "k97..."
  }
]

Step 7: Agent B (Bob) recalls the memory

{
  "query": "project brief MVP",
  "namespace": "global",
  "limit": 3
}

Response includes the memory Agent A (Alice) stored — with semantic search ranking.

Step 8: Agent B (Bob) marks the message as read

{
  "receiptIds": ["k97..."]
}

Done. Two agents, shared memory, real messaging, read receipts. No file hacks. No polling. No duct tape.

What just happened

1. One Convex deployment serves as the shared backend for both agents
2. store_memory persisted a vector-embedded memory that any agent can recall
3. send_message delivered a message from Alice to Bob with a receipt
4. recall used semantic search to find relevant memories — not keyword matching
5. mark_as_read confirmed Bob processed the message

Next steps

• Add tasks so agents can assign work to each other
• Set up recurring tasks for cron-based automation
• Read the full architecture to understand orchestrators, instances, and namespaces
• Browse all 82 MCP tools