MCP
The same 44 operations, as tools an assistant can call.
The Model Context Protocol is how an AI assistant reaches a system without anyone writing code. Posttar serves one at https://mcp.posttar.com, and every tool there is an endpoint of this API: same operations, same names, same refusals. Nothing exists in one and not in the other.
Connecting
In an assistant that takes a connector (Claude, ChatGPT), paste the address and authorize in the browser. The assistant registers itself, you pick which workspace it may reach, and no key is copied anywhere.
In a client where you can set headers (an editor, a terminal), use the same address with an API key in Authorization: Bearer. It is the key from Settings → API & MCP, the same one this API takes.
Claude Code, with a key:
claude mcp add --transport http posttar https://mcp.posttar.com \
--header "Authorization: Bearer $POSTTAR_KEY"What the server does, and what it does not
| Answer | |
|---|---|
| Transport | Streamable HTTP. One POST, one JSON reply. |
| Session | None. Every call carries its own credential, and nothing is remembered between them. |
| Tools | One per endpoint. GET /posts is get_posts, POST /posts is post_posts. |
| Resources and prompts | Not offered. The server declares only tools. |
| Server-initiated stream | GET answers 405: there is nothing the server needs to say on its own. |
A tool that refuses answers with the refusal, not with a protocol error: the same code and message this API returns, so the assistant can read what went wrong and fix the call instead of giving up.
Authorization
OAuth 2.1 with PKCE, discovery included: https://mcp.posttar.com/.well-known/oauth-protected-resource says who authorizes, and the authorization server publishes its own metadata. Clients register themselves. Access tokens last an hour and refresh tokens rotate, so a refresh token used twice ends the whole connection.
A connection belongs to one workspace, like a key does. You can see and disconnect what you authorized in Settings → API & MCP.