> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getsly.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP (Model Context Protocol)

> Anthropic's protocol for exposing tools and resources to LLM agents.

MCP is Anthropic's Model Context Protocol — a standard for how LLM agents (Claude, GPT, Gemini) discover and invoke tools exposed by external servers. Sly provides MCP on two surfaces:

1. **The `@sly_ai/mcp-server` package** — runs locally (stdio transport) and gives Claude Desktop, Cursor, or Windsurf direct access to Sly capabilities
2. **Remote MCP over HTTP** — Sly-hosted MCP server at `/mcp` for agents that need remote tool access

## Local MCP server (Claude Desktop, Cursor, etc.)

Install:

```bash theme={null}
npx @sly_ai/mcp-server install
```

This adds Sly to your MCP client's config. Restart the client; Sly tools appear in the tool catalog.

Tools exposed:

* `sly_create_account`, `sly_list_accounts`, `sly_get_account`
* `sly_create_agent`, `sly_list_agents`
* `sly_create_wallet`, `sly_get_wallet_balance`, `sly_fund_wallet`
* `sly_create_transfer`, `sly_get_transfer_status`
* `sly_create_stream`, `sly_withdraw_stream`
* `sly_ap2_create_mandate`, `sly_ap2_execute_mandate`
* `sly_ucp_create_token`, `sly_ucp_settle`
* `sly_x402_pay`, `sly_x402_verify`
* `sly_a2a_send_task`, `sly_a2a_list_tasks`, `sly_a2a_complete_task`

See the [MCP tool catalog](/agents/mcp-tool-catalog) for the full list with schemas.

## Remote MCP over HTTP

For agent runtimes that don't run locally but can speak MCP, Sly exposes `/mcp` directly:

```bash theme={null}
# JSON-RPC over HTTP
curl -X POST https://api.getsly.ai/mcp \
  -H "Authorization: Bearer agent_..." \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/list",
    "id": 1
  }'
```

Returns the tool catalog. Invoke a tool:

```bash theme={null}
curl -X POST https://api.getsly.ai/mcp \
  -H "Authorization: Bearer agent_..." \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/call",
    "params": {
      "name": "sly_get_wallet_balance",
      "arguments": { "wallet_id": "wal_..." }
    },
    "id": 2
  }'
```

Supports three transports:

* **POST `/mcp`** — JSON-RPC request/response (default)
* **GET `/mcp`** — Server-Sent Events streaming (for long-running tools)
* **DELETE `/mcp/:session_id`** — Close a session

## Authentication

MCP over HTTP requires one of Sly's auth methods. Recommended:

* **Agent token** (`agent_*`) — simplest
* **Ed25519 session** (`sess_*`) — production

API keys work but grant too broad a scope for typical agent use cases.

## Tool discovery

Every registered MCP tool has:

* A stable `name` (e.g. `sly_create_transfer`)
* An `inputSchema` (JSON Schema)
* A `description` optimized for LLM consumption
* Permission requirements (what KYA tier / scope can invoke it)

Agents receive only the tools their credentials allow them to invoke.

## Scanner MCP (separate server)

Sly also ships an MCP server for the [Scanner service](/api-reference) — a demand-intelligence tool for merchants and agents.

```bash theme={null}
npx @sly_ai/scanner-mcp install
```

Exposes 19 tools for scanning merchants, analyzing protocol adoption, generating readiness reports, and running agent shopping tests. See the [scanner docs](/api-reference) for detail.

## When to use MCP

* You're building an LLM agent that needs to transact (Claude, GPT, local models)
* You want **drop-in tool access** rather than calling the REST API manually
* Your runtime already speaks MCP (Claude Desktop, Cursor, Windsurf, custom agent frameworks)
* You want **automatic permission filtering** — MCP server only exposes tools the agent can actually call

For non-LLM integrations (backend services, CI, operational scripts), use the [REST API](/api-reference) directly.

## MCP tool catalog

The full list of every tool in `@sly_ai/mcp-server` and scanner MCP is in the [MCP catalog](/agents/mcp-tool-catalog). Each entry includes input schema, example invocation, and permission requirements.