Agent Kit exposes MCP tools over POST /mcp (JSON-RPC tools/list and tools/call). Do not rely on a fixed tool list in documentation — names and count change as Agent Kit grows.
Discover on every connect. Call tools/list (or open the dashboard Tools page) for the live catalog: each entry includes name, description, and inputSchema. Wire your LLM or backend from that response.

​
Authentication

POST https://agent-kit.abstraxn.com/mcp
Authorization: <application-api-key-or-agent-api-key>
Content-Type: application/json
Key typeagent_id in arguments
Application API key (dashboard Overview)Required when multiple agents exist
Per-agent API key (createAgent → agent.apiKey)Optional; must match that agent if provided

​
Tool categories (OAuth scopes)

OAuth and consent use scopes, not individual tool names. New tools are included when they fit the same category.
ScopeWhat it covers
mcp:tools:readWallet/chain reads and x402 catalog discovery
mcp:tools:transferUnsigned transfer intents — your backend signs and broadcasts
mcp:tools:paidFixed-price and external x402 HTTP calls
After OAuth or API-key connect, call tools/list to see which names are available for your granted scopes.

​
Calling a tool

  1. tools/list → pick a name and read inputSchema.
  2. tools/call with params.name and params.arguments (JSON object matching the schema).
  3. For application API keys, pass agent_id in arguments when the schema requires an agent (typical for wallet actions).
{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "<from-tools-list>",
    "arguments": {
      "agent_id": "AGENT_UUID"
    }
  }
}
Some tools return payment required (-32402) on the first call. That is expected:
  1. Parse accepts / payment metadata from the error or tool result.
  2. On your backend, sign paymentPayload with the agent accessKey.
  3. Retry tools/call with paymentPayload at the top level of params (sibling to name / arguments).
Unsigned transfer intents are different: Agent Kit returns an unsigned transaction; your backend signs with accessKey and broadcasts — it does not use the x402 retry flow. Details: x402 payments.

​
Example: list tools

APP_KEY="<dashboard-api-key>"
MCP="https://agent-kit.abstraxn.com/mcp"

curl -s -X POST "$MCP" \
  -H "Authorization: $APP_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq .
Use a name from result.tools[] for any tools/call example in Smoke test.

​
LLM integration pattern

The sample backend API (NestJS reference):
  1. Sets MCP_SERVER_URL + MCP_SERVER_AUTH_TOKEN (= application API key).
  2. On startup (and after reconnect), runs initialize + tools/list — map whatever the server returns.
  3. Maps discovered tools to OpenAI (or other LLM) function definitions.
  4. During chat, the LLM may call tools/call; results are fed back into the model.
  5. Paid tools trigger x402 signing in AgentSigningService.
Your frontend only talks to your JWT API — never to MCP directly.

​
Next steps