Quick start

Three levels of integration - from exploring in the portal to connecting your own AI app

Meko has three levels of integration. Each gives you more control over how your agents use memory, knowledge, and conversation history. Start at level 1 and work up, or skip ahead.

Level	What you connect	What you get
1. WorkbenchLM	Our chat agent, in the portal	Simplest. See memory and knowledge work before installing anything.
2. Your AI client	Cursor, Claude Code, Claude Desktop, Codex, Kiro, OpenCode, VS Code	Daily driver. Use your own coding tools against your datapacks.
3. Your AI app	Your own agent or framework	Full control. Build Meko into the apps you ship. Under development.

Before you start, sign up or sign in to the Meko portal. You land on the Welcome to Meko page with a default datapack already created for you.

New accounts start on the Free tier. See Meko pricing for plan details and limits.

Note

Meko itself is not an AI agent, it's the data layer (shared knowledge, collective memory, decision traces) that your agents connect to. WorkbenchLM is an embedded chat agent so you can try Meko in the portal without installing anything; it's not the daily-use surface.

Try Meko in WorkbenchLM

The simplest way to see Meko in action. WorkbenchLM runs in the portal against your default datapack, with sample knowledge loaded and live memory traces visible as you chat. Nothing to install.

In the portal sidebar, open WorkbenchLM.
Ask about the sample knowledge, or share something about your project for Meko to remember.
Watch the Memory and Decision traces panels update as the conversation moves.

See WorkbenchLM for what's loaded, when to use it versus a real coding agent, and the limits of the embedded experience.

Connect your AI client

External AI clients (Cursor, Claude Code, Claude Desktop, Codex, Kiro, OpenCode, VS Code, and so on) connect to Meko through one shared MCP endpoint:

https://mcp.mekodata.ai/mcp

You don't need a different URL per datapack. Your agent picks the datapack at call time by passing a datapack_id (or datapack name) into MCP tool calls. API keys and the MCP connection are scoped to your account, not to a single datapack. Once connected, your agent can see and act on every datapack you own.

There are three ways to connect. Pick the one that fits your workflow.

Recommended for Cursor, Claude Code, Claude Desktop, Codex, Kiro, and OpenCode. Requires Node.js 20.18.1 or newer.

The installer registers the Meko MCP server and installs an agent skill that teaches your agent when to use Meko tools, including saving conversations, capturing memories, and searching knowledge. In clients that support them, it also configures session hooks.

On Claude Desktop, the installer builds two .skill files and you must import them yourself. See Claude Desktop.

Automatic capture in installer 2.5.0

Claude Code, Cursor, Codex, Kiro, and OpenCode keep failed exchanges in a local outbox and retry them. Installer 2.5.0 also repairs stuck capture about every 30 minutes. Run npx @yugabytedb/meko-mcp@latest --capture-cleanup fix to repair now. Kiro and OpenCode don't capture tool calls or hidden reasoning. Claude Desktop has no session hooks, so its automatic capture still depends on the model.

Automatic capture writes to your pinned datapack by default, or to your account's default datapack if you haven't pinned one. OpenCode CLI and TUI use the account's default datapack and don't read the pin. See Automatic capture and datapack pinning.

In a terminal, run:
```
npx @yugabytedb/meko-mcp@latest --client <client> --url "https://mcp.mekodata.ai/mcp"
```
Replace <client> with claude-code, claude-desktop, cursor, codex, kiro, or opencode.

With MEKO_API_KEY unset, the installer opens your browser. Sign in the same way you sign in to the Meko portal. It creates an API key and stores it for that client. The key is listed under Settings > API Keys with a name like Meko installer (<computer>) <time>. It counts toward your 10 active keys and expires after one year.

OpenCode defaults to user scope, which applies to every project. To install it only in the current project, run the command from the project root and add --scope project. See OpenCode CLI and TUI.

On a fresh install, you can omit --url. The installer offers https://mcp.mekodata.ai/mcp, then asks whether to sign in with the browser or paste a key. If a client already has a Meko endpoint, that endpoint is the default. With --yes, clients that point at different endpoints stop the install until you pass --url. An interactive install warns and asks you for the URL, with no default.

To paste a key you already have, add --api-key <paste-your-api-key>. If MEKO_API_KEY is already set, the installer uses that key and does not open a browser. The Connect dialog on a datapack still shows that form. Over SSH, add --no-browser and open the URL the installer prints. A script or CI job has no terminal, so pass --api-key or set MEKO_API_KEY, and add --yes. --yes in a terminal still signs in through the browser.
Restart your client if the new MCP server doesn't show up. OpenCode always requires a restart after installation.

The portal also has a Manual MCP Install tab for other clients, which walks through running nvm commands and adding the skills package step by step.

See @yugabytedb/meko-mcp on npm to see what the installer does under the hood.

For Claude Desktop and other clients with a custom-connector flow.

If your client supports adding a custom remote MCP connector, you can connect Meko by URL with no API key and no installer. The client opens a browser for OAuth; your Meko portal login is the identity.

In your client, add a custom MCP connector pointing at https://mcp.mekodata.ai/mcp.
Complete the OAuth handshake when the browser opens.
The MCP server returns its tool list to the client automatically.

OAuth doesn't work for every client

Clients that complete OAuth through a local http://127.0.0.1 callback currently fail with an HTML 403 Forbidden. That covers command-line clients and some IDEs - it's about the callback the client registers, not whether it has a window. Use an API key for those. Hosted connectors with an https callback, such as Claude Desktop's, are unaffected.

With OAuth alone, your agent interacts with Meko on demand: you need to mention "Meko" in your prompt for the agent to call the tools. For example, "save this to Meko", "search my Meko knowledge base". The agent won't proactively save conversations or capture memories unless you ask.

To get automatic behavior, use the Automated install instead, or write your own skills to teach your agent when to call Meko tools.

For any MCP-compatible client - no npx, no installer.

Add the Meko MCP server directly to your client's config file with your API key. This gives you the same on-demand MCP tools as OAuth, but authenticated with a long-lived key instead of a browser handshake.

Go to Settings > API Keys and click Generate API key. Name it for the client that will use it (for example, laptop/cursor). The full key is shown exactly once: click Copy or Download .txt before closing.

Edit your client's MCP config to add the Meko server. For example, in Cursor's mcp.json:

{
  "mcpServers": {
    "meko": {
      "url": "https://mcp.mekodata.ai/mcp",
      "headers": {
        "Authorization": "Bearer <paste-your-api-key>",
        "User-Agent": "meko-mcp-installer/1.0"
      }
    }
  }
}

Keep the Bearer prefix - Authorization is the only header Meko reads for authentication. Keep User-Agent too; Meko rejects requests without one. See Choose an authentication method.

Restart your client.

Like OAuth, this is on demand: your agent calls Meko tools when you ask. For per-client configuration paths and examples (Claude Desktop, Claude Code, Codex, VS Code), see Connect to an AI agent.

For API key generation, revocation, and limits, see API keys.

Verify the connection

Ask your agent to list your datapacks.
```
List my datapacks in Meko.
```
It should call the datapack_list tool and return the names you see in the portal. If it doesn't, see Troubleshooting.

If you can see more than 1,000 datapacks, datapack_list returns an error that tells the agent to use datapack_search, which returns one page at a time.

Connect your AI app

Under development. Programmatic integration patterns for agent frameworks (LangChain, LlamaIndex, custom Python or TypeScript agents) are coming. See Connect programmatically for the current state and what's planned.

Explore Meko

Add knowledge to your datapack and create memories.