Glasser MCP Server

Glasser plugins

One key, no contract, pay per call, across paid APIs bought wholesale.

Official agent plugins for Glasser. This repository is a distribution package: configuration plus markdown. It ships no executable code and no install script.

Works with Cursor, Claude Code, Grok Build, and any client that reads the Agent Plugins format.

Install

Cursor

Search glasser under Customize → Plugins, or open the listing on the Cursor Marketplace. The first call opens your browser to sign in.

Claude Code

/plugin marketplace add glasser-ai/plugins
/plugin install glasser@glasser

Any supported agent

npx plugins add glasser-ai/plugins

Detects every agent installed on the machine and installs to all of them.

Grok Build

/marketplace inside Grok Build once the plugin is listed. Until then, add this repository under [[marketplace.sources]] in ~/.grok/config.toml; the Grok Build docs have the syntax.

MCP only

Any MCP client that signs in with OAuth needs just the URL:

https://api.glasser.ai/mcp

Clients that cannot open a browser can send a Key instead — client-by-client snippets: https://glasser.ai/docs/mcp-server

Authentication

Nothing to create or paste. The first tool call opens your browser: sign in to Glasser, choose the Workspace whose balance the agent may spend, and click Allow. That mints a Key named MCP · <client> for that Workspace — it is listed at https://app.glasser.ai/keys like any other Key, and revoking it there disconnects the client. Disconnecting inside the client does not revoke the Key; the next sign-in reuses it.

The optional CLI authenticates on its own with glasser login (browser device authorization). That stores a Key for the CLI only.

What you get

MCP tools — search, inspect, run, runs_get, runs_list, runs_stop, balance. Only run creates a Run and charges the Workspace balance; the other tools read the catalog, past Runs and the Balance. Every Run reports its own charge.

Skill — the workflow the agent follows: search the catalog, inspect an Endpoint's Price and charge clauses before running, run, then report the result and the charge. It uses the MCP tools by default and the CLI for results too large for the context window, for scripting and for CI.

Rule (Cursor only) — always-on spending guardrails.

Optional CLI — npm install -g @glasser-ai/cli. Same Key, same Workspace, same Balance. Docs: https://glasser.ai/docs/cli

Network endpoints and credentials

For reviewers and anyone auditing what this plugin reaches.

Nothing in this repository executes. The plugin sends no telemetry of its own. The service behind these endpoints is governed by Glasser's terms of service and privacy policy.

Repository layout

.cursor-plugin/marketplace.json   Cursor catalog (one entry)
.claude-plugin/marketplace.json   Claude Code catalog (one entry)
plugins/glasser/
├── plugin.json                   Agent Plugins manifest
├── .cursor-plugin/plugin.json    Cursor manifest (carries rules/)
├── .claude-plugin/plugin.json    Claude Code / Grok Build manifest
├── openclaw.plugin.json          OpenClaw manifest (ClawHub requires it to publish)
├── mcp.json                      MCP server (read by Cursor, Agent Plugins)
├── .mcp.json                     identical copy (read by Claude Code, Grok Build)
├── skills/glasser/SKILL.md
├── rules/glasser-spending.mdc
├── assets/logo.svg               1:1, transparent (see plugins/glasser/README.md)
└── assets/icon.png               the same mark, rasterised for ClawHub
scripts/                          validators and the version bump

Releasing: see DISTRIBUTION.md.

License

MIT — see LICENSE. The Glasser service has its own terms.