Glasser Research

Use Glasser to find and run paid third-party data API endpoints through one account. Glasser provides access and billing; the selected provider supplies the response. Use the provider's name when attributing data.

This is a data collection skill. Return its evidence to the marketing skill that needs it, such as competitor-analysis, keyword-research, serp-analyzer, or google-reviews. Keep that skill's analysis method and output format.

Source selection

Follow an explicit tool choice from the user. Otherwise use sources in this order:

1. A suitable free tool already available in the agent.
2. A working integration or API key the user already configured.
3. Glasser for the specific remaining data gap.

Do not activate Glasser only because an API appears in its catalog. Use it when the task needs live data that the current environment cannot retrieve. A failed search can mean that the topic has no results; distinguish that from missing access before offering a paid fallback.

Setup

Check the CLI and authentication before planning a paid run:

glasser --version
glasser balance

If the CLI is missing and the user asks to configure Glasser, install it with Node.js 22 or later:

npm install -g @glasser-ai/cli@latest

For an interactive session, run glasser login. Keep the command active while the user approves the matching code in the browser. Relay the fallback URL and code when the browser does not open. Never ask the user to paste a key into chat. After login, run glasser balance again in the environment that will make calls.

For unattended environments, the user can configure GLASSER_API_KEY through a secret manager. Do not write keys to project files, reports, command arguments, or chat. If the environment key overrides a saved login and fails, fix or remove that override instead of repeating login.

If the host exposes Glasser MCP tools, use balance, search, inspect, run, and runs_get instead. Follow the current MCP setup instructions for client setup.

Research workflow

1. Define the evidence needed, target market, result volume, and freshness window. Avoid broad collection when a small result set answers the question.
2. Run glasser search -q "<capability>". This searches the endpoint catalog, not the web or social platform. Endpoint descriptions are not research evidence.
3. Compare relevant providers, then run glasser inspect -p <provider> -e <endpoint>. Read the current price, charge clauses, input schema, volume controls, and run mode. Never guess fields from a provider's direct API documentation because Glasser's contract can differ.
4. Tell the user the endpoint, provider, request size, and expected cost. Get approval before spending unless the user already authorized this exact scope or an adequate task budget. Do not make speculative, repeated, or bulk calls.
5. Write the inspected input as JSON to a task-specific temporary file. Use -f rather than interpolating user text into a shell command. Choose an unused output file so existing research is not overwritten.
6. Run the endpoint. Use --wait for asynchronous work and -o for large output:

glasser run -p <provider> -e <endpoint> -f <input.json> --wait -o <output.json>

7. Read the provider response. COMPLETED means the provider answered; it does not guarantee a useful result. Missing fields stay unknown. Do not infer a zero value or invent a metric.
8. Return the evidence to the calling marketing workflow. Report the provider, query and market, retrieval date, result limitations, final charge, and the private Glasser Run URL. Cite public source URLs from the provider response; the Run URL is an audit record for workspace members, not a public citation.

Common marketing data

These examples were inspected on 2026-09-12. Search and inspect again before a run because coverage, schemas, and prices can change.

SERP example

After inspecting serper /search, write an input file such as:

{"q":"project management software","gl":"us","hl":"en","num":10}

Use returned organic positions, titles, snippets, and public links. A SERP result does not establish keyword volume, organic difficulty, traffic, or conversion.

Community language example

After inspecting scrapecreators /v1/reddit/search, write an input file such as:

{"query":"project management software complaints","timeframe":"month","sort":"top","filter":"posts"}

Preserve the post date, author, subreddit, URL, and engagement fields only when the provider returns them. Treat posts as opinions from their authors, not facts about the market. Retrieve comments only for selected posts and only within the approved budget.

Recovery and limits

• For QUEUED or RUNNING, wait on the existing run with glasser runs get -r <runId> --wait; do not submit the same job again.
• On an ambiguous timeout or dropped connection, retry once with the same Idempotency-Key. JSON mode and MCP run require an explicit UUID.
• On rate limiting, wait for the returned retry delay and retry the same request once. Do not loop.
• On insufficient balance, stop and ask the user to top up. Do not reduce the request in a way that changes the agreed research question without saying so.
• Money values are exact decimal strings. Do not sum them using binary floats.
• Send only the query and public identifiers needed for the approved task. Do not send private CRM exports, local files, or personal data unless the user clearly authorized that data and the selected endpoint requires it.