Cursor Enterprise can send Grok Bot and Cloud Agent activity directly to Weave over OpenTelemetry. A Cursor workspace administrator and permission to manage Weave connections and API keys are required.
This integration accepts only cursor.surface=grok_bot and cursor.surface=cloud_agent. Desktop, CLI, SDK, Bugbot, unknown and missing surfaces are excluded, even when Cursor sends them in the same export. Your existing Cursor plugin continues collecting local activity, and the Cursor Admin API remains the source of billed usage and cost.
Cloud Agents launched through an SDK are included when their resource has cursor.surface=cloud_agent. Filtering follows the surface, not the entrypoint; only the separate sdk surface is excluded.
- In Weave, open Connect Cursor → Cursor Enterprise plan → Grok Bot and Cloud Agents.
- Enter a connection name and your numeric Cursor team ID, then create the connection. Weave issues a fresh shared organization Cursor key for this destination. Personal keys cannot authenticate a cloud export.
- Copy the displayed destination base URL and key. The key is shown only once. The URL uses your Weave instance’s existing
/ingest collector endpoint, including for self-hosted installations. Native exports and the Cursor plugin share this endpoint.
- In Cursor’s Team Settings → OpenTelemetry Export, configure that base URL and set the
X-Weave-Token header to the key. Cursor appends /v1/logs and /v1/metrics; do not append either signal path to the base URL.
- Run Cursor’s destination test. Then generate Grok Bot or Cloud Agent activity and refresh the connection’s delivery status in Weave.
The destination must be publicly reachable over HTTPS and accept OTLP/HTTP binary protobuf. Cursor supports one export destination. If your team already exports to another collector, configure that collector to forward both logs and metrics to Weave, preserving resource attributes and either the X-Weave-Token header or each resource’s weave.token attribute. The Weave collector carries the header into that resource attribute before batching. Mixed native and plugin resource groups are routed separately.
Understand coverage
Weave stores known event metadata, delta metric points, and useful exported content: user prompts, assistant replies, tool arguments/results, commands, browser details, repository/commit context, and named skill/hook/plugin and cloud action details. Content export requires Cursor’s team opt-in and destination toggles. Weave’s organization and member content controls apply before publication and again before storage. Unknown fields, future categories, and unsupported events or metric kinds are ignored without rejecting valid sibling records.
Bodies retain separate observed, provider-truncated, and Weave-redacted indicators. Tool content remains text even when truncation makes its JSON incomplete. Missing turns and unmatched messages remain independent records; a persistent Bot ID does not imply that separate content and action turns match. The connection activity list remains metadata-only. Transcript access and prompt analysis are delivered separately; storing content alone does not provide those reports or session cost.
Prompt and assistant content follows the Cursor retention family, including Grok Bot exports. Temporary content releases after its real turn becomes idle, or independently by event when no turn was supplied. Content wipes retain diagnostic metadata and fence delayed queue messages. Replayed events never restore removed content.
Enable Action Recording in Cursor to receive Grok Bot action logs. Without it, other supported events and metrics may still arrive. Missing events cannot tell Weave whether Action Recording or a content toggle is disabled.
The configuration page separates:
- Configured: the team has a connection and a bound key.
- Last accepted in-scope export: Weave accepted a nonempty Grok Bot or Cloud Agent batch for processing.
- Last processed logs / metrics: the respective signal was successfully stored. Either signal can arrive independently.
An empty destination test checks connectivity to the collector. The collector carries credentials on individual records, so an empty test cannot verify key validity, the bound export connection or the Cursor team. Generate cloud activity to verify these checks and data delivery. Empty tests and excluded-surface traffic do not update in-scope delivery timestamps. Recent activity lists only event names, surfaces, timestamps and outcomes; it contains no member identity or conversation content.
Logs may be replayed; Weave deduplicates by event identity. Metrics may have delivery gaps, and any exported cost is an estimate rather than an invoice total. Cursor does not export historical activity from before the destination was enabled.
Rotate or disconnect
Choose Create replacement key, copy the new key into Cursor and run its destination test. Both keys remain bound during this overlap. After verifying delivery, unbind the old key. Unbinding removes its access to this export without revoking the shared Cursor key for other uses.
Disconnect export prevents new ingestion for that connection and preserves existing observations. A team ID cannot be edited on an existing connection; create a new connection for a different team.
Troubleshoot delivery
- Authentication failures: check the header name, whether the key is still bound, and whether its shared key has been revoked.
- Team mismatch: use the Cursor team ID configured in Weave. Payload attributes cannot override the credential’s organization.
- No processed data after a passing test: use Grok Bot or a Cloud Agent. Local Cursor activity is intentionally filtered. Check that your collector forwards both signals and preserves
cursor.surface.
- Rejected large batches: requests allow at most 6 MiB after decompression and 4,096 resources. Split batches at your collector; Weave never silently truncates them.
See Cursor’s export setup and wire reference for provider settings and event coverage.
Conversation and action details
Select a recent activity event to open its retained conversation and action details.
The viewer shows independent provider events, including assistant-only messages and
unpaired tool arguments/results, with the exact supplied correlation IDs. Cursor
truncation and privacy redaction are labeled; missing exports cannot be reconstructed.
The UI window covers 15 days on either side of the selected event.
Content access requires member-report visibility and the prompt-content grant for
other members; your own retained content uses the existing own-content exception.
Unassigned content requires organization-report visibility and the prompt grant.
Connection-settings permission alone does not grant transcript access. Current
category and retention settings apply to every read, and reporting anonymization
withholds native transcripts.
GET /agent-trajectories?source=cursor_otlp&start=<RFC3339>&end=<RFC3339> returns the same native events in
cursor_otlp.nodes. The organization API key requires agent_trajectories scope.
Native pages default to 50 events and are capped at 100 with native_limit; use
cursor_otlp.next_cursor as native_after with the same explicit start, end,
and source. Windows are capped at 31 days. The sealed cursor freezes the event
population while privacy changes still apply to each page. Existing Claude Code
and local Cursor exports keep their own trajectory limits and response fields. Last modified on November 13, 2026