Appearance
Telemetry vs Enterprise
Two Agent capture pathways. Pick the one that matches what your engineers can actually consent to today — you can flip later. Telemetry captures metadata only and works whether your team is on consumer or enterprise plans. Enterprise captures the full prompt corpus and unlocks fluency scoring and policy enforcement, on contracts that cover it.
The Cloud Proxy is a separate shipped setup path for network proxy routing and subject metering. It does not, by itself, turn proxied traffic into Enterprise prompt-body capture. Use the Flowstate Agent when you need the capture modes described below.
Agent Capture Modes
| Telemetry | Enterprise | |
|---|---|---|
| What's captured | Request metadata: host, model, tokens, status, duration | Full request + response, including prompts |
| Account types | Consumer + enterprise plans both work | Designed for Claude / ChatGPT / Copilot Enterprise |
| Attribution | Per person, team, project, cost centre | Per person, team, project, cost centre |
| Prompt-quality scoring | — | ✓ |
| Performance / price prediction | ✓ | ✓ |
| Provider-billing reconciliation | ✓ | ✓ |
| Policy enforcement | - | Yes (optional) |
| Audit trail of prompts | - | Yes |
| Data residency control | Metadata only | Full residency choice |
| Storage footprint | Low | Higher (prompt + response retained) |
When to pick which
Telemetry is the right answer when:
- Your engineers are on a mix of consumer accounts (personal Claude Pro, personal ChatGPT Plus) and you can't compel an account switch.
- You want spend attribution and seat insight without taking responsibility for the prompt corpus.
- You're starting with governance and will add enforcement later.
Enterprise is the right answer when:
- You're on Claude for Enterprise, ChatGPT Enterprise, GitHub Copilot Business / Enterprise, Cursor Business, or comparable contracts with a real admin surface and a data-handling agreement.
- You need prompt-quality scoring at the person + team level.
- You intend to enforce policy: provider allow-lists, model bans, cost-centre routing, monthly caps.
- You want the audit trail that satisfies a SOC 2 or ISO 27001 reviewer.
If you're between the two, deploy Telemetry first. The agent is the same binary in either pathway — flipping the capture mode is a setting change, not a redeploy.
What the agent sees in each pathway
| Wire data | Telemetry | Enterprise |
|---|---|---|
| Hostname | Yes | Yes |
| HTTP method, path, status | Yes | Yes |
| Request + response headers used for attribution | Yes | Yes |
| Token counts (input / output / cache / cache-read) | Yes | Yes |
| Model identifier | Yes | Yes |
| Request body (prompts, attachments, tool calls) | No | Yes |
| Response body (assistant text, structured output, SSE frames) | No | Yes |
Headers are used for attribution in both modes. Credentials are not shown in customer-facing session views.
Switching between pathways
The capture mode is a single setting at Settings → AI → Agent Policy → Capture mode. Changes propagate to the fleet on the agent's next poll (within an hour by default).
There's no schema migration involved. Telemetry-only data captured under "Telemetry" remains queryable; the enriched Enterprise data starts flowing from the moment the setting changes.
Next
- Install the agent: macOS · Windows · Linux
- Use the Cloud Proxy for fleet proxy routing and subject metering when a full Agent rollout is not the right first step.
- Choose the policy that matches your pathway: Agent policy