127.0.0.1, with nothing leaving your machine.
It’s the same loop as hosted Glassray, packaged as one process you run yourself: great for trying Glassray before connecting production, debugging locally, working offline, or running in CI.
Setup is one paste: on an empty store, glassray-coach start (and the dashboard’s empty state) hands you an onboarding prompt for your coding agent - paste it into Claude Code, Codex, or Copilot and it discovers your flows and rules from your code, wires tracing, and verifies the first trace lands. Then you just run your agent.
Coach binds
127.0.0.1 only and every route enforces a loopback guard. Your traces stay on your machine - no account, no upload.What you get
Live trace viewer
Traces stream in as your agent runs. Open any one for a full span waterfall, inputs/outputs/attributes, and one-click Replay of an LLM step.
Deviation discovery
An LLM judge reads your traces and clusters the recurring ways your agent misbehaves - the silent failures ordinary monitoring misses.
Durable flows
Name your agent’s behaviours once - a selector query and/or a plain-language rule, or inferred straight from your agent’s code (
flows discover) - and Coach classifies every new trace into them in the background, across sessions.Self-rerunning evals
Turn any deviation - or a single trace - into a pass/fail check scoped to its flow. When enough new traces land, the eval reruns on its own and flags regressions.
Self-healing fixes
One click turns a deviation into a concrete fix - instructions you paste into your coding agent (what to grep for, which files, across prompt / tools / guardrails / code).
Flows as code
Flows, run recipes, and your code root live in
glassray.yaml - round-tripped with pull / push and gated in CI with check, so your rule suite travels with the repo.A/B experiments
Compare a baseline against a candidate - say a cheaper model - over the same rule suite, and read the verdicts side by side.
Runs on your model
Analysis runs against your local Claude subscription (zero-config), a metered API key, or a deterministic offline
mock.The loop
- See - every run lands live; open a trace to inspect the waterfall or replay an LLM call with an edited prompt.
- Find - discovery clusters recurring failures into deviation types, each with a plain-language rule.
- Scope - name the behaviour as a durable flow (a selector query and/or a plain-language rule); Coach classifies new traffic into it in the background, so the scope stays current across sessions.
- Fix - click Generate fix on a deviation and Coach writes a concrete fix as instructions for your coding agent (Claude Code / Cursor): what to grep for, the likely files, and the ordered edits across prompt, tools, guardrails, and code.
- Verify - save the deviation’s rule as an eval bound to its flow; as fresh traces classify in, the eval reruns automatically, the pass rate climbs, and anything that breaks a formerly-passing case is flagged as a regression. Mark the deviation resolved to close the loop.
Coach vs. hosted Glassray
Coach - local
One process on your machine. Your traces never leave. No account. Best for trying Glassray, local debugging, offline, and CI.
Glassray - hosted
A continuous loop over your production traffic that learns intent from your code and proposes code-level fixes as pull requests. Join the alpha.
Next steps
Quickstart
Get Coach running and see your first trace in about three minutes.
Send traces
Instrument your agent with the SDK or any OTLP exporter.
Find & fix deviations
Run discovery, save evals, and prove your fixes with a worked example.
CLI & coding agents
Every command, plus the agent skill (Claude Code, Codex, Copilot) that drives the whole loop.