MCP server

How it works#

The Model Context Protocol lets an AI client call tools a server exposes. PostStack's server wraps the same REST API the TypeScript SDK calls: every tool call becomes one or more ordinary API requests made with your API key, so the same permissions, rate limits, suppression checks and sending rules apply as if your own code had made them.

Both serve the identical set of tools, prompts and resources. Create the key under API keys in the dashboard, and read Permissions before choosing its permission level.

Connect the hosted server#

Point any MCP client that speaks Streamable HTTP and lets you set a request header at https://api.poststack.dev/mcp, with the header Authorization: Bearer sk_live_…. The server is stateless: each request builds a fresh server, so there is no session to keep alive or resume. In Claude Code:

claude mcp add --transport http poststack https://api.poststack.dev/mcp \
  --header "Authorization: Bearer sk_live_..."

Requests to the endpoint are limited per API key: 120 a minute on paid plans, 60 on the Free plan, on top of the limits of the API endpoints each tool calls. The current ceiling is in the X-RateLimit-Limit response header. Responses also carry X-Poststack-Mcp: v1.

Run it locally#

The @poststack.dev/mcp package runs as a stdio server that your client starts on demand, so there is nothing to install ahead of time.

1. Add the server to your client

   Claude DesktopClaude CodeCursor

   Copy

   {
     "mcpServers": {
       "poststack": {
         "command": "npx",
         "args": ["-y", "@poststack.dev/mcp"],
         "env": { "POSTSTACK_API_KEY": "sk_live_..." }
       }
     }
   }

   Claude Desktop reads ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows. Cursor reads ~/.cursor/mcp.json, or .cursor/mcp.json in a project.

2. Restart the client

   Claude Desktop and Cursor read the file at startup. Once connected the client lists the PostStack tools; ask it to “list my sending domains” to confirm the key works.

npx @poststack.dev/mcp --print-config <target> prints a snippet for claude-desktop, cursor or claude-code, filled with POSTSTACK_API_KEY if it is set. The claude-desktop and cursor output is the whole file shown above, ready to save; the claude-code output is the claude mcp add command.

POSTSTACK_API_KEYstringrequired

The key every tool call authenticates with. Without it the process prints POSTSTACK_API_KEY environment variable is required and exits with status 1.

POSTSTACK_BASE_URLstringdefault https://api.poststack.dev

API origin the tools call.

POSTSTACK_MCP_TRACE"1"

Writes one JSON line per tool call (timestamp, tool, ok, duration, error code) to stderr, never to stdout, so it cannot corrupt the protocol stream.

Permissions#

The server can do exactly what its API key can do. Tools that need more than the key allows return the API's 403 as a tool error instead of acting.

Tools#

117 tools in 16 groups. Arguments marked ? are optional; the client shows the agent each argument's type and description. Every tool returns a short text summary plus the data as structuredContent. When the API rejects a call the tool returns isError: true with structuredContent of { statusCode, code } and the API's error message, so the agent can tell a validation error from a missing permission.

Prompts#

Prompts are playbooks the client offers you by name (in Claude, from the prompt picker). Each tells the agent which tools to call in which order. tone is formal, friendly or casual and defaults to friendly.

Resources#

Read-only JSON documents the client can attach to the conversation before the agent acts.

Usage analytics#

Every tool call made through the hosted endpoint is recorded and shown on the MCP analytics page of the dashboard: total calls, error rate, p50 and p95 latency, top tools, hourly activity and recent errors, over the last 24 hours, 7 days or 30 days. Calls made through the local stdio server are not recorded.

Each record holds the tool name, duration, success or failure, error code, transport and the API key used, plus a hash of the arguments used to group identical calls. The arguments themselves are never stored. The hash is not cryptographic, so do not rely on it to hide low-entropy values such as an email address. Records older than 30 days are deleted.

Troubleshooting#

Next steps