X Bookmarks MCP: Connect Saved Posts to Codex and Claude Code

The phrase Model Context Protocol can sound more abstract than it is.

"Model" is the AI system that writes, reasons, or plans. In this article, that usually means the model inside Codex, Claude Code, Claude, or another agent app.

"Context" is the extra information the model needs. A model can answer general questions from its training, but it does not automatically know your saved X posts, your project notes, your favorite design examples, or the libraries you bookmarked last month. That outside information is context.

"Protocol" literally means a set of rules. In software, a protocol is an agreed format for communication. MCP gives AI apps a common way to ask external tools for information or actions.

So an X bookmarks MCP means an AI app can ask a saved-X connector for relevant posts using a shared tool format.

MCP has a few technical words that are worth defining before they become confusing.

The client is the app where the agent runs. Codex, Claude Code, Claude Desktop, and Cursor can be MCP clients in the right context. "Client" just means the software asking for something.

The server is the connector that exposes actions. In socialmemory's case, the public package is @popoastier/socialmemory-mcp, a local MCP connector for saved X likes and bookmarks. "Server" does not always mean a big cloud machine. A local MCP server can run on your own computer and talk to the client through a controlled channel.

A tool is one action the server offers. Examples include searching saved X posts, reading one saved post, checking sync status, adding a note, merging tags, or creating a result-set link.

This split matters because a good MCP setup exposes specific tools with clear names and narrow jobs.

Generic web search starts broad. Your saved X memory starts narrow. It contains things that already passed through your attention. That makes it especially useful for agents, because the agent can search inside your taste and history instead of searching the whole public web from scratch.

For example, "Search my saved X posts for onboarding examples, then suggest patterns that fit this app" is more personal than asking for generic onboarding advice.

Retrieval means getting the right thing back later. Saving is easy. Retrieval is hard.

You may remember that you saved a post about a SQLite library, but not the author's name. You may remember a landing-page teardown, but not the exact phrase in the post. You may remember saving several posts about "AI agents with memory," but not whether you liked them, bookmarked them, or saw them inside a thread.

Native X search can help when you remember exact words, an author, or a time period. A saved-X library helps when you remember the idea but not the exact post. MCP matters when an agent should do that search during work.

The normal Codex setup path is one copied setup prompt from Agent Access. The verified command shape is:

In plain language:

The setup flow should install the bundled Codex plugin and MCP server config, open browser-approved login, store the credential privately, run setup checks, and show status. A credential is a private proof that the connector is allowed to access your socialmemory account. It should not be pasted into chat.

The setup check matters. A successful setup is not only "the browser opened." It should also prove that server access works and the archive status can be checked.

Claude Code has its own MCP setup path. The socialmemory web app should provide a separate setup prompt for Claude Code, using Claude Code's MCP command format plus browser-approved socialmemory login.

The article should not imply the Codex command is also the Claude Code command. The practical copy can say: Codex has a one-command setup path, and Claude Code has a separate generated setup path in Agent Access.

Agent Access should not create a second, lower-quality archive. Chrome extension sync and agent sync should produce the same saved-post quality over time.

Normal refresh should be incremental. A full resync should be treated as an advanced recovery action, not the everyday "Sync Now" behavior.

Before starting a new project, ask the agent to search saved posts for the domain, technology, or product pattern.

Example prompt:

> Search my saved X posts for anything about onboarding, activation, and first-run UX. Group the results into patterns, examples, warnings, and posts worth opening.

That gives the agent a personal research pack before it starts making suggestions. For a coding project, Codex can combine the repo context with the saved-post context. For Claude Code, the same idea applies inside a local development workflow.

Builders often save posts about libraries, framework patterns, command-line tools, database tips, browser APIs, or debugging threads.

Example prompt:

> Search saved X posts for SQLite, local-first, and sync libraries. Return the posts that look relevant to this app, then explain which ones are worth reading first.

The agent should retrieve candidates, summarize why each may matter, and let you decide.

Saved X posts often contain screenshots, landing pages, app flows, teardown threads, and product positioning ideas. An MCP search can turn that into a working swipe file.

Example prompt:

> Find saved posts about pricing pages and explain the repeated patterns. Create a result-set link with the strongest examples.

That result-set link lets the human inspect the same source material.

Founders often save advice that is hard to categorize in the moment: launch stories, pricing experiments, positioning examples, growth loops, hiring lessons, and investor notes.

Example prompt:

> Search my saved X posts for launch advice, waitlists, and founder-led distribution. Pull out concrete tactics, then separate strong examples from vague advice.

That last instruction tells the agent not to treat every saved post as equally true.

After a useful search, the agent can help attach light structure.

Example prompt:

> For the saved posts you found about AI agent memory, suggest tags. Add only the tags I approve, and create a result-set link for this research pack.

The agent can suggest structure, but sensitive or broad changes should remain explicit.

"Least privilege" means giving software only the access it needs for the job. If the job is searching saved posts, the connector should not need broad permissions to modify unrelated data.

For saved X memory, a safe default is read-first:

Least privilege also reduces damage if something goes wrong. A narrow connector is easier to reason about than a tool that can do everything.

A token is a secret text key that proves access. A credential is the broader private access proof stored by the setup flow. The important rule is simple: do not paste secrets into normal agent chat.

The socialmemory setup should prefer browser-approved login and local credential storage. On macOS, the CLI can store the credential in Keychain. Keychain is Apple's private credential storage system. If a visible raw token exists, it should remain an advanced fallback, not the normal consumer setup path.

Good article wording:

Risky article wording:

Socialmemory sync depends on the local X/browser session when collecting saved posts from the user's machine. A good setup check should confirm that the browser is signed into X and readable before starting sync work. That prevents misleading failed sync records caused by a local browser problem.

This is especially important for first sync. The first sync should be full, newest-first, and progressive so the library can fill over time. After that, normal sync should be incremental.