MCP Server

Zafeguard runs a Model Context Protocol (MCP) server alongside the SDK API. Point any MCP-compatible client — Claude Desktop, Claude Code, Cursor, or your own — at it and your AI assistant can:

• Inspect workflows, components, and runs.
• Execute workflows and components on demand.
• Build workflows directly on the canvas — drop in components, position them, wire data ports and sequence edges, attach dynamic variables.

The server speaks the MCP Streamable HTTP transport over a single endpoint.

Endpoint

POST https://api.zafeguard.com/v1/mcp

Enabling MCP for a workspace

MCP is disabled by default on every workspace. To turn it on:

1. Open Workspace Settings → scroll to MCP — AI agent access.
2. Flip the toggle to on. A token like zmcp_abc12345… is generated and shown once.
3. Copy the token immediately and paste it into your MCP client config. The raw value is shown once and never again — if you lose it, click Rotate token to issue a new one (the old one stops working instantly).

You can disable MCP at any time without losing the token; re-enabling later reactivates the same value. Requests to /v1/mcp from a disabled workspace are rejected with 401 Unauthorized — MCP is disabled for this workspace.

Authentication

MCP uses a workspace-specific MCP token (zmcp_…), not the SDK API key (ws_…). The token is dedicated to MCP traffic and can be rotated independently of your SDK keys.

Send:

There is no Ed25519 request signing — MCP clients can't sign. Rate limits, monthly/weekly usage caps, and credit-debt rules still apply to MCP traffic.

Tools

All tools are scoped to the workspace that owns the MCP token. Tools that reference a workflowId verify the workflow belongs to your workspace before doing anything.

Help

Read tools

Execution tools

Workflow-builder tools

These let the agent construct workflows visually — same effect as drag-and-drop in the canvas editor.

Versioning tools

Agents can iterate on the draft and read history. Publishing a draft to live and restoring earlier versions are dashboard-only actions — not available via MCP.

All tools return their payload as a single text content block containing pretty-printed JSON.

Setup — Claude Desktop

Claude Desktop loads MCP servers from claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Add a zafeguard entry under mcpServers:

{
  "mcpServers": {
    "zafeguard": {
      "transport": "http",
      "url": "https://api.zafeguard.com/v1/mcp",
      "headers": {
        "x-mcp-token": "zmcp_your_workspace_mcp_token_here"
      }
    }
  }
}

Restart Claude Desktop. The Zafeguard tools should appear in the slash-command menu (/tools).

Setup — Claude Code

Add the server with the claude mcp add command:

claude mcp add --transport http zafeguard https://api.zafeguard.com/v1/mcp \
  --header "x-mcp-token=zmcp_your_workspace_mcp_token_here"

Or edit ~/.claude/mcp_servers.json directly with the same shape as the Claude Desktop config above.

Verify the connection:

claude mcp list
# zafeguard   http  https://api.zafeguard.com/v1/mcp   connected

Setup — generic HTTP client

Any client that can issue JSON-RPC over HTTP can talk to the server. Each request is a JSON-RPC 2.0 envelope. Example: list tools.

curl -X POST https://api.zafeguard.com/v1/mcp \
  -H "x-mcp-token: zmcp_..." \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/list"
  }'

Call a tool:

curl -X POST https://api.zafeguard.com/v1/mcp \
  -H "x-mcp-token: zmcp_..." \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc": "2.0",
    "id": 2,
    "method": "tools/call",
    "params": {
      "name": "execute_workflow",
      "arguments": {
        "workflowId": "3f7a1b2c-...",
        "input": { "userId": "u_123", "amount": "100" }
      }
    }
  }'

Example prompts (Claude)

Once the server is connected, you can ask Claude natural-language questions:

"List the workflows in my Zafeguard workspace."

"Run the daily-treasury-sweep workflow with input { "vault": "0xabc..." } and tell me when it completes."

"Get the status of run 9e4c7b1a-.... If any stage failed, summarise the error."

"Build me a simple workflow that takes an EVM address as input, fetches its USDC balance, and emits the value. Place the trigger at [0,0], the balance check at [240,0], and wire the address through."

Claude picks the right tools, calls them with arguments inferred from your prompt, and renders responses.

Errors

The server returns standard JSON-RPC error envelopes:

A 429 from a rate-limit hit includes a Retry-After header (seconds).

Security model

• The token is dedicated to MCP — rotating it doesn't affect SDK API keys, and vice versa.
• Only the SHA-256 hash and an AES-GCM-encrypted copy of the token are persisted. The plaintext only ever lives in your MCP client config.
• Every tool closes over the workspace ID resolved from the token — there's no way for a tool argument to "escape" the workspace. Workflow-builder tools additionally verify the target workflow belongs to the same workspace before mutating it.
• The endpoint is stateless: each request instantiates a fresh McpServer and tears it down on response close. No cross-request state.
• The SDK's Ed25519 signing path remains the more secure option for server-to-server integrations — use it whenever you control both ends.

Related

• Authentication — full API-key + signature flow used by the SDK.
• Execute Component — the same operations exposed as HTTP endpoints.
• Execute Workflow — workflow triggering reference.