nc://glasser

Glasser: 1,000+ paid data endpoints behind one key, as seven tools. A Nodecode rewrite of glasser-ai/hermes-plugin-glasser by glasser-ai.

Glasser is a data broker: one key covers endpoints from providers such as Apollo, People Data Labs, Hunter, Ahrefs, Semrush, DataForSEO, Serper, Exa, ScrapeCreators and RentCast, and each call is charged the endpoint's published price. The original is a manifest that points a coding agent at Glasser's remote MCP server. This cell is that server, reached over MCP's streamable HTTP with the cell's own client, so it needs no Node bridge and no npx.

Get it

Tell the agent "add the glasser cell", or run nodecode add nc://glasser.

Set it up

Make a key at app.glasser.ai/keys, then put it in the glasser section of config.jsonc (or /settings):

{
  "glasser": {
    "api_key": "…",                                  // or leave it out and export GLASSER_API_KEY
    "endpoint": "https://api.glasser.ai/mcp",       // the public server, the default
    "timeout_seconds": 60                            // one call's ceiling, a run included
  }
}

api_key_env names an environment variable instead, and GLASSER_API_KEY is read either way, so a key exported in a shell needs no section at all. The key can be revoked at app.glasser.ai/keys; a section with no key refuses at start and says where to make one. /glasser answers which key is in use and against which endpoint.

Use it

The seven tools keep the original's names and arguments, under the cell's prefix:

Every run carries an idempotency key: a retry with the same key returns the first run instead of paying twice, and a new key is a new charge. When the model leaves the key out, the cell makes a UUID for it; a caller that means to retry passes its own.

Through eval, the same seven are one verb:

(glasser:call "balance")
(glasser:call "runs_list" (nlk:json-object "limit" 5))

While the cell runs, (help :glasser) answers the working agreement the original ships as a skill — search, then inspect, then run; name the price before a speculative, looped or bulk run; poll a QUEUED or RUNNING run; read charge_usd on every run; treat insufficient_balance and rate_limited as answers, not as errors to retry blind.

Not ported

• The browser sign-in. The original's bridge, mcp-remote, opens glasser.ai in a browser and caches the token it gets. A cell takes a key from its section instead — the same key the sign-in resolves to, listed at app.glasser.ai/keys — so nothing here holds an OAuth token or refreshes one.
• The Node bridge itself. mcp.json runs npx -y mcp-remote@0.14.3; this cell speaks the same MCP endpoint over HTTP directly.
• The skill as a skill. Its guidance is (help :glasser) here, one manual topic rather than a file loaded through skills_list.

Gaps

• The cell declares each tool's arguments as the original's README and skill describe them (query, endpoint, input, idempotency_key, run_id, limit); it does not read the server's own tools/list schemas before registering, so a server-side rename would show as the server's own refusal rather than as a repaired argument list.
• Every call opens its own MCP session, initializes it, and leaves it for the server to expire: a cell call is one exchange, and no state is kept across calls.
• The release's MCP client already reaches this same server by config (mcp.servers.glasser), with its own OAuth sign-in. This cell is for installing Glasser as one thing, with its tools named, its key in one section, and its spending rules on the manual.

Test

.github/test-cell.sh tools/nodecode-glasser

Every HTTP exchange is a stubbed nlk:http, so nothing reaches Glasser or the network: the lifecycle, the key's two ways in, /glasser, and each answer shape — a balance, a run with a generated key, a refusal the server reports, a refused key. A green run ends 12 glasser tests, 0 failed.

MIT licensed. Glasser's own guidance and manifest were written by glasser-ai; see NOTICE.