Skip to main content
Ask Glassray is the in-product way to explore your traces conversationally. Instead of writing filters, you describe what you’re looking for in plain English and work with the result set across turns. In the dashboard it’s the docked chat panel you open with ⌘G (the ⌘K command palette can also hand a query off to it). It’s built on three primitives that operate on a working set - the set of traces you carry from one turn to the next.

Find

Describe what you’re looking for in plain English and get back a curated result set - trace ids, a one-line summary, and a reason per match. If the request is ambiguous, it asks a clarifying question instead of guessing.

Analyze

Ask a question about the current set (“what do these have in common?”) and get a written analysis plus the outliers that don’t fit.

Refine

Narrow or grow the set with an instruction (“keep only the ones that errored”, “add similar ones”) and get the new set back.

​
A working set, carried across turns

The three primitives compose. A typical session looks like: find a set of traces, analyze what they share, then refine down to the ones that matter. Each step takes the working set as input and returns the next one, so you can pivot without starting over.
Ask Glassray runs metered LLM work and can take a few seconds - especially find, which searches your whole corpus.

​
Same engine, two surfaces

The exact same three primitives are exposed as tools on the product MCP server - find_traces, analyze_traces, and refine_traces. Whether you ask from the dashboard panel or from your coding agent over MCP, it’s the same engine over your org’s traces.
Connect over MCP to drive find → analyze → refine from Claude Code, Cursor, or Claude Desktop.