See how your AI tools are actually used.
Version-aware telemetry for AI coding environments. Which skills, which versions, which agents, which cost — from Forge compiler manifest to Grafana. Your data never leaves your infrastructure.
The enterprise visibility gap
No idea which skills are deployed
You sync 20 skills to 50 developers. Which version is each developer running? Are all teams on the latest? No way to know.
No cost attribution
Langfuse shows total LLM cost. But which skill caused it? Did skill v2.1 cost 40% more than v2.0? Nobody can answer this.
Trust problem
Any observability tool that phones home to a third party is dead on arrival in enterprise. Data must stay in your infra.
Prometheus model for AI assistants
Crucible provides the collector. Company provides the destination. Zero data touches Crucible servers.
The closed feedback loop
Forge manages what configs/skills agents get. Ember measures how those configs/skills are actually used. Together: configure → measure → optimize.
Enterprise value
For Engineering Managers
"How much are we spending on AI per team? Which teams are actually using the tools we paid for? What's the ROI?"
For Platform Teams
"Which skills are used? Which should we deprecate? Are developers accepting or reverting AI suggestions?"
For Compliance & Security
"Audit trail of AI usage. Proof that guardrails are applied. Data stays in our infrastructure, zero third-party exposure."
Built on top of existing telemetry
Ember's data starts with the Forge compiler manifest — a JSON file emitted on every build that lists every skill, version, hash, agent target, and token estimate. No hooks needed. Then it grows with integrations.
Tier 1 — Forge Manifest
Which skills, which versions, which agents, per project. Version drift detection across developers. Token budget utilization per agent.
Tier 2 — Langfuse / Helicone
Correlate manifest data with actual LLM usage. Answer: 'skill X v2.1 costs 40% more tokens than v2.0.' Aggregate across tools.
Tier 3 — Agent OTEL
Ingest native agent telemetry (Claude Code OTEL, Cline logs). Confirm which skills were actually activated, not just deployed.
Each tier adds a new data dimension without replacing the previous one. Start with manifest analytics, grow into full observability.
What you'll see
SKILL INVOCATIONS — LAST 7 DAYS
TOKEN USAGE BY TEAM
ASSISTANT DISTRIBUTION
SKILL VERSION ADOPTION
Mock data — illustrates what Ember dashboards will provide. Push to Grafana, DataDog, or any OTLP-compatible backend.
How Ember fits the landscape
| Tool | What it does | Ember's relationship |
|---|---|---|
| Langfuse | LLM trace observability | Ember imports Langfuse data + correlates with skill versions for cost attribution |
| Helicone | LLM proxy with logging and caching | Ember enriches Helicone logs with skill-version context |
| Claude Code OTEL | Native skill_activated events, token traces | Ember ingests + adds versioning, cross-tool normalization |
| DataDog LLM Obs | Enterprise LLM monitoring | Ember pushes OTLP metrics with skill-version dimensions to DataDog |
| Grafana | Open-source dashboarding | Ember ships a pre-built Grafana dashboard pack — import and go |
Ember doesn't compete with any of these. It's the unified collector that covers all AI assistants and pushes to your choice of backend.