Skip to main content
MCP lets your AI tool use Glasser directly. Connect your account to search for data, check prices, and retrieve results in your conversations. You do not need to install the Glasser CLI.

​
Choose your AI tool

Claude Code

Add Glasser from your terminal and sign in from Claude Code.

Codex

Connect local Codex clients with a shared MCP configuration.

Claude.ai

Connect Claude on the web or desktop, including Cowork.

ChatGPT

Connect Glasser as a plugin and use paid data in ChatGPT.

Manus

Connect Glasser and use paid data in your research and reports.

Cursor

Add Glasser to Cursor and enable it for your agent.

OpenCode

Add Glasser to your OpenCode config and sign in from the terminal.

VS Code

Add Glasser to GitHub Copilot Chat in VS Code.

Other clients

Connect by server URL or use an existing Glasser Key.

​
Already know your client’s setup?

Use this server URL and choose OAuth sign-in:
Sign in to Glasser, choose the Workspace whose balance the client can use, and select Allow. You can create an account and Workspace during sign-in. If your client cannot use OAuth, follow Other clients to connect with a Key.

​
Check the connection

Ask your AI tool:
A balance response confirms that the connection works, even if the available balance is zero. The check is free. Add funds in the Glasser console before a paid Run.

​
Manage access

Browser sign-in creates a Key named MCP · <client> in your Workspace. To remove the client’s access, revoke that Key on the Keys page. Disconnecting or removing Glasser in the client alone does not revoke the Key.

​
Endpoint

Use Streamable HTTP. Authenticate through browser sign-in or send a Glasser Key in the Authorization: Bearer header.
The server is stateless and returns JSON responses. Each request includes either a Key or an OAuth access token that resolves to a Workspace Key.OAuth clients discover the authorization server from resource_metadata in the WWW-Authenticate header of a 401 response. They need only the server URL to start this flow.

​
Tools

Glasser exposes seven tools. The Agent uses search to discover Endpoints, then inspect and run to retrieve data. Each Endpoint does not require a separate MCP tool.

​
Rules that keep spending safe

The server provides these instructions to the Agent, as the Skill does for CLI agents:
  • Use existing connections first. Use Glasser where your own Keys and integrations do not cover the request.
  • Check before spending. Inspect the Price, start with small requests, and get approval before speculative, repeated, or bulk Runs.
  • Retry with the same key. Keep the idempotency_key and execution parameters to retrieve the original Run without a second charge.
  • Read status and charge separately. Check provider_response and charge_usd, even for a COMPLETED Run.
A replayed MCP Run includes replayed: true. Amounts in charge_usd are exact USD decimal strings. A Provider 404 can be a COMPLETED Run charged under the Endpoint’s NO_RESULT rule.Provider errors remain in the Run’s provider_response.http_status and provider_response.error. Check these fields even when the tool result has no isError flag. Use runs_get for the full error body; runs_list returns summaries.Actionable refusals, such as insufficient_balance, rate_limited, and Policy restrictions, return tool results with isError. They are not protocol errors. For rate limits, read retry_after_ms before retrying.

​
When to prefer the CLI

For local agents such as Claude Code, Codex, and Cursor, the Skill plus the CLI is the recommended path. The Skill guides setup and carries the working rules. For Claude Chat and Cowork, follow the connector guide.