Developer Tools

Implement Zep with agents

Install the Build with Zep plugin so Claude Code, Codex, and Cursor build on Zep with the current docs and best practices.

Coding agents build on Zep faster when they understand how Zep works and can read the current documentation as they write code. The Build with Zep plugin gives Claude Code, Codex, and Cursor both: a skill that encodes how to design, build, and evaluate a Zep integration, and the Zep documentation MCP server for real-time search over the docs. One plugin installs in all three ecosystems.

The plugin is for building with Zep in your own codebase. You’ll still need a Zep project and API key — start with the quick start guide.

What the plugin includes

The plugin bundles two things that work together:

  • The building-with-zep skill — the decision-and-workflow layer for building on Zep: scoping graphs, ingesting data, retrieving context, and evaluating whether Zep delivers your use case. Your agent invokes it automatically when you write or design Zep integration code. It indexes the docs rather than duplicating them, so its guidance stays current.
  • The Zep documentation MCP server (zep-docs) — real-time search and whole-page access over Zep’s documentation, served at https://docs-mcp.getzep.com/mcp (remote HTTP, no API key). The skill queries it for exact, current details — method names, parameters, limits — and prefers reading a full page over a search snippet.

The plugin is open source — the skill, its MCP configuration, and the Claude Code / Codex / Cursor marketplace catalogs live in the getzep/building-with-zep-plugin repository.

Install

From your terminal (not inside a Claude Code session), add the plugin repository as a marketplace and install the plugin:

$claude plugin marketplace add getzep/building-with-zep-plugin
$claude plugin install building-with-zep@building-with-zep

Keep the plugin up to date

The zep-docs MCP server is remote, so your agent always searches the current documentation. The skill ships as a file on disk, so it goes stale: new guidance, workflows, and corrections reach you only when the plugin updates. Claude Code needs updating switched on; Codex does it by default. Without that, your agent keeps building against whatever the skill said the day you installed it. The plugin’s changelog lists what each release changed.

Claude Code enables auto-update by default only for Anthropic’s own marketplaces. building-with-zep is third-party, so updates stay off until you turn them on: run /plugin, select Marketplaces, choose building-with-zep, then select Enable auto-update.

Claude Code then refreshes the marketplace and updates the plugin shortly after each session starts, with a random delay of up to ten minutes. Your running session keeps the version it launched with. If the plugin changed, Claude Code prompts you to run /reload-plugins; otherwise the new version loads the next time you start it.

To update once without enabling auto-update, refresh the marketplace and then the plugin:

$claude plugin marketplace update building-with-zep
$claude plugin update building-with-zep@building-with-zep

Administrators rolling the plugin out to a team can skip the per-user toggle by setting "autoUpdate": true on the marketplace’s extraKnownMarketplaces entry in managed settings.

If you set DISABLE_AUTOUPDATER to manage Claude Code’s own updates yourself, it disables plugin updates too. Set FORCE_AUTOUPDATE_PLUGINS=1 alongside it to keep plugin auto-update running.

How your agent uses it

Once installed, the skill activates on its own when you work on Zep: adding memory to an agent, ingesting chat or business data, searching the graph, choosing between user and standalone graphs, or debugging retrieval. As it works, the agent reads the current documentation through the zep-docs MCP server instead of relying on its training data, so its code and recommendations track the live docs — and when the two disagree, the docs win.

Just the documentation MCP server

In a client that supports MCP but not skills, set up the Zep documentation MCP server on its own. Your agent gets the current documentation, without the design and evaluation guidance the skill carries.