Appearance
AI Telemetry
AI Telemetry captures usage data from AI coding tools across your engineering organization — token consumption, costs, session activity, and CapEx classification — without changing how engineers work.
What It Does
- Capture every AI request from a developer's machine, three ways — OTel from instrumented tools, an on-device proxy daemon, or the hosted relay at
ai.flowstate.inc - Token and cost tracking per engineer, team, and project
- CapEx / R&D classification for tax relief reporting
- Author guidelines, see activity, manage an attention list of subjects who need a check-in
- Zero-config for engineers (CLI installer handles setup, MDM artefacts ship with the binary)
Capture modes — pick one or mix
| Mode | What you install | Coverage | Best for |
|---|---|---|---|
| OTel via CLI | flowstate-telemetry install | The 9 supported tools | Richest signal (file paths, tool calls, git context) |
| On-device proxy | flowstate-telemetry proxy install | Every AI request from the machine | Bypass-proof fleet rollouts. MDM-deployable. |
| Hosted relay | None — point a base URL at ai.flowstate.inc | Whichever tools let you set a base URL | Locked-down machines, contractors, CI runners |
See the capture modes overview for detail.
Supported tools (OTel mode)
| Tool | Collection Method | Prompt Capture |
|---|---|---|
| Claude Code | OTel (native) | Opt-in (--prompts) |
| GitHub Copilot | OTel (VS Code) | Opt-in (--prompts) |
| Cursor | Editor hooks | Always on |
| Windsurf | Editor hooks | Always on |
| Gemini CLI | OTel (native) | Opt-in (--prompts) |
| Codex CLI | OTel (native) | Opt-in (--prompts) |
| Aider | OTel | Always on |
| Qwen Code | OTel (native) | Opt-in (--prompts) |
| OpenCode | OTel (native) | Always on |
The on-device proxy and the hosted relay are tool-agnostic — they intercept the API call to the upstream provider, so they work with anything that hits Anthropic / OpenAI / Gemini / Copilot regardless of which tool issued it.
What Gets Tracked
For every AI coding session:
- Token counts — input, output, and cache-read tokens per model
- Cost — estimated USD cost based on provider pricing
- Session metadata — duration, tool calls, files modified, lines added/removed
- Git context — branch, repository, commit messages (when available)
- Classification — project attribution and CapEx eligibility (automated)
WARNING
For Claude Code, Copilot Chat, Gemini CLI, Codex CLI, and Qwen Code, prompt capture is opt-in via the --prompts flag. For Cursor, Windsurf, Aider, and OpenCode, prompt content is captured by default when telemetry is configured.
How It Works
- Engineers install
flowstate-telemetryCLI (OTel mode),flowstate-telemetry proxy install(on-device proxy mode), or point a tool's base URL atai.flowstate.inc(hosted relay mode) - AI tool usage reaches the Flowstate telemetry service via whichever path was configured
- Sessions are aggregated, classified (project + category + CapEx eligibility), and stored
- Data appears in the Flowstate AI Workforce dashboard, where admins can author guidelines, watch the activity log, and manage an attention list for subjects with significant unmatched-session spend
Next Steps
- Engineers (any one tool): OTel setup guide — get telemetry running in under 5 minutes
- Fleet rollout: On-device proxy — install via Jamf / Intune / Kandji
- Locked-down machines: Hosted relay — no install, just a base URL
- Admins: Governance — guidelines, activity log, attention list
- Tool-specific config: Claude Code, GitHub Copilot, Cursor, and more