Wire an AI client to the 1Claw MCP server

The 1Claw MCP server gives an AI client tools to list, get, and manage vault secrets. You can use the hosted server at mcp.1claw.co or run a local stdio process.

1. 1

   Option A: Local stdio (recommended for development). Add the 1claw MCP server to your client config. Only ONECLAW_AGENT_API_KEY is required. Agent ID and vault are auto-discovered. Paste your real ocv_ key from agent creation in place of the placeholder.

   json Copy

   {
     "mcpServers": {
       "1claw": {
         "command": "npx",
         "args": ["-y", "@1claw/mcp"],
         "env": {
           "ONECLAW_AGENT_API_KEY": "ocv_paste_your_agent_api_key_here"
         }
       }
     }
   }

2. 2

   Option B: Hosted server. First get a JWT by calling the agent-token endpoint. Replace the placeholders with your real agent_id and ocv_ key from agent creation.

   bash Copy

   export AGENT_ID="paste-your-agent-uuid-here"
   export AGENT_KEY="ocv_paste_your_agent_key_here"
   
   TOKEN=$(curl -s -X POST https://api.1claw.co/v1/auth/agent-token \
     -H "Content-Type: application/json" \
     -d "{\"agent_id\":\"$AGENT_ID\",\"api_key\":\"$AGENT_KEY\"}" | jq -r '.access_token')
   
   export VAULT_ID="paste-your-vault-uuid-here"

3. 3

   Use that token in a hosted MCP config. Claude Desktop, Cursor, and Claude Code all use this same shape. Replace $TOKEN and $VAULT_ID with the values from the previous step.

   json Copy

   {
     "mcpServers": {
       "1claw": {
         "url": "https://mcp.1claw.co/mcp",
         "headers": {
           "Authorization": "Bearer $TOKEN",
           "X-Vault-ID": "$VAULT_ID"
         }
       }
     }
   }

4. 4

   Restart the client, then ask it to list your secrets.

   text Copy

   List the secrets in my 1claw vault.

Your AI client now calls list_secrets against your vault and returns the secret paths and metadata.

What this gives you, and what it does not. MCP hands an assistant a set of tools; it does not hand it judgment, and it does not change who is allowed to do what.

• The MCP server enforces nothing. Every call is authorised by the vault API against the agent's policies, so connecting an assistant grants exactly the access the agent already had.
• Tool descriptions enter the model's context. Anything you install contributes text the model reads as guidance, which makes an installed server's metadata part of your attack surface.
• Output inspection is a backstop, not a boundary. Known secret values are redacted before reaching the model, and a credential in an unfamiliar format is not.
• The stdio server holds your agent key. It comes from the client config on your machine, so it is only as protected as that file.