Data access & privacy

The levels

You can run the whole loop at the basic level: with no content flags, SightLift still measures your work and proves capabilities forward from live use.

1. Basic setup (start here)

Follow Live telemetry setup to point Claude Code / Cowork at your ingest endpoint, and add the SightLift tool (MCP connector) in your tool's settings (your admin can push it org-wide). That alone runs the loop — no content access needed.

2. Better classification (optional)

export OTEL_LOG_USER_PROMPTS=1
export OTEL_LOG_ASSISTANT_RESPONSES=1

Adds real prompt/response text so SightLift classifies your work more precisely than metadata alone.

3. Automation harvesting (optional)

To let SightLift build proven, deterministic automations from work you already do in Claude Code, add tool content. (These are Claude Code settings. Cowork is configured in the admin console, not with environment variables.)

# the command / script
export OTEL_LOG_TOOL_DETAILS=1

# the tool output (requires tracing)
export CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1
export OTEL_TRACES_EXPORTER=otlp
export OTEL_LOG_TOOL_CONTENT=1

With these on, SightLift can see the exact command and its result, reproduce it against your real examples, and only then offer it back as a verified automation.

4. claude.ai chats (optional, Enterprise)

To include claude.ai chat work, create a Compliance Access Key in claude.ai with the read:compliance_user_data scope and add it in SightLift's integration settings. SightLift reads claude.ai chats, their messages and your user list with it. Claude Code and Cowork reach SightLift through steps 1–3.

Security & privacy

Can we start with no content at all? Yes. Basic setup plus the SightLift connector runs the whole loop — SightLift proves capabilities forward from live use, so nothing in steps 2–4 is required to get value.

What exactly do the content options expose? Step 2: your prompt and Claude's response text. Step 3: the commands your tools run and their output. Step 4: your claude.ai chat messages and tool activity (the key itself reaches more; see step 4). Each is a separate switch you control and can turn off at any time.

What does the SightLift hook send? If you install the optional prompt hook (the proactive suggestion trigger), it sends each prompt's text to SightLift to check for a matching capability — even if your telemetry content options are off. It suggests something only on a strong match, and never shares the prompt beyond your organization. If you'd rather not send prompts, skip the hook: the SightLift tool still works passively (your assistant can find capabilities by name), just without proactive suggestions.

Where does content go, and how is it protected? Content is scoped to your organization only — never shared across tenants. Live telemetry, tool content included, is kept for about 24 months and then expires automatically. Data you bring in by upload or an API sync is kept until you delete the dataset it arrived in (below). Where a capability was built from particular conversations, their tool inputs and outputs are retained so we can keep re-proving it as your work changes. All of it is tenant-scoped and subject to the return, deletion and erasure terms in your DPA (§9 and §11).

Can a manager see how much sensitive work one person does? No. SightLift never reports how much of one person's work falls in a sensitive, protected or not-yet-classified category. What it does report is a breach of your company's policy: sensitive work outside the person's recorded clearance, a match of a rule your admins added, or work on an account that isn't on your roster. Those alerts name the person to the people who review governance in your company, so they can follow it up, and they are not included in exports. My AI shows you how many of your own conversations are protected. In one person's topic breakdown and in a list of their conversations, the ones you may not read are shown as restricted, with no title or category: you can see how many there are, not what they are.

What about deletion / GDPR? Two paths, and it is worth knowing which one you need.

Deleting a whole dataset is self-serve: under Configuration → Data & sources → Datasets, an admin deletes a dataset and every conversation it brought in goes with it. This is the operative path in the product today.

Erasing one person is a right-to-erasure request you make to us. It is keyed to the person rather than to a single conversation, and it removes them from the conversation records, the dashboard data, the raw telemetry events and the historical score rosters in one pass. There is no button for it in the app yet, so send the request to privacy@sightlift.ai or your account team. Two things to expect: any capability built from that work keeps its de-identified proof snapshot, which carries no personal data and is treated as aggregated data under your DPA (§10); and if the data came in by manual upload, the original export file is kept as the source audit trail and is not rewritten, so say if you need that purged too.

Do we have to enable everything to get started? No. Most customers start at step 1 + the connector, add step 2 for accuracy, and enable step 3/4 for specific teams once their security review is comfortable.

Need help choosing a level for your risk profile? Contact your SightLift account team.