From an answer to an action.

Ask a coding agent to explain a change and it can produce a summary. Ask it to publish that summary and it needs another capability: a tool that can perform the action.

The Model Context Protocol, or MCP, provides a common way for AI applications to connect to servers that expose capabilities. An application can discover a server’s tools and call them with structured inputs. The protocol also supports resources and prompts. These are connection mechanics; the actual capability comes from the server.

For capx café, the relevant capability is publishing to X. The agent supplies text to tools such as preview and post_now. The same writing context can therefore support both drafting and the next step, without requiring you to transfer the draft into a separate posting interface.

Keep three parts of the workflow distinct.

  1. The agent interprets your request. It uses the context you provide to reason about the work and write a draft.
  2. The MCP client connects the tools. capx café runs as a local stdio MCP server that your compatible agent can launch.
  3. The capx café service controls the send. It applies publishing checks and accesses the X token on the server.

“Local MCP server” and “server-side token” describe different parts of this setup. The local process exposes the tools to your agent. The hosted or self-hosted service keeps the publishing credential in its vault. Installing the local package alone does not give you a connected X account.

Follow one request all the way through.

Suppose you ask your agent to turn a release into an X post. First, it reads the release material you authorized. Next, it drafts the wording. You can then ask it to preview that wording through capx café, review the result, and authorize a send.

The service receives the post text routed through its tools; it does not gain general access to the repository simply because your agent can read it. Your agent’s own access and privacy settings remain relevant.

Try this with your agent

Explain which capx café tools are available in this session. Separate tools that only inspect or preview from tools that publish or change a schedule. Do not publish, connect an account, or create a schedule.

Choose tools by their boundaries, too.

When you evaluate an agent integration, ask what it can read, what it can change, where credentials live, and how you can inspect the result. A common protocol does not make every connected tool equally appropriate for every task.

capx café keeps the X access token on the server and puts its deterministic guardrail in the publishing path. Your agent still has the ability to request guarded posts through a session handle. That ability deserves deliberate use.

The practical opportunity is a shorter path from useful context to a useful action. Begin with a small workflow you understand, inspect the tool results, and expand only when you are comfortable with what each step does.

Sources & further reading