Payload capture mode
You choose how much tool call input and output endpoints send. The setting applies to every endpoint and agent.
Even with Omitted, endpoints still send redacted command and request summaries. For calls Merlin flags, they also send redacted user-request and tool-history context.
Endpoints report each tool call’s decision, mode, tool, endpoint and timing with every option, including Omitted.
Endpoints apply changes to new tool calls after receiving the setting. Full payloads cannot recover earlier content. Choose the setting before you need that content. Decisions alone are enough to see what a policy would block. Choose Full payloads if you need to see arguments, file contents or destinations when you investigate.
Kontext redacts secrets and credential values before storing payloads. It also limits payload size. Redaction looks for credentials, so other confidential content in files or command output can still get through. Choose the setting based on the data your agents handle. In Explore, Kontext marks redacted payloads with Secrets redacted.
Resource scan
Kontext scans agent configuration on every endpoint. The scan reads MCP servers, plugins, permission settings and whether credentials exist. The report includes credential presence but excludes secret values. The dashboard uses the scan for the Elevated access and Resources cards on Home. Kontext enables the scan by default. When you turn it off, the dashboard asks you to confirm before deleting stored resource reports. Tool call records and payload capture stay unchanged. Runkontext report on an endpoint to see its last report. Use kontext report --json for machine-readable output.
Usage data for Cost
For Cost, endpoints read agents’ local session files. They send model names, token counts, timing and links to each model request’s tool calls. Endpoints exclude prompt text, model reasoning, tool arguments and tool output from usage records. Endpoints send usage records even with Omitted payload capture. See how cost is calculated.Risk detection
Kontext runs Risk detectors on the endpoint and records findings with the decision. For flagged tool calls, Kontext runs a hosted AI assessment of the redacted tool call and nearby session context. The endpoint decides without waiting for the assessment. The local guardrail model is off by default. You can turn it on or off per endpoint with Guardrail LLM on Home. See detection.Local records and removal
Each endpoint makes and records decisions locally. The endpoint exports records to your workspace in the background. The dashboard may show them after a short delay.kontext setup --uninstall preserves the endpoint’s local identity and records and prints their paths. The command leaves the Homebrew binary installed. Your workspace keeps the endpoint’s recorded activity. To delete hosted records, follow your workspace’s agreed retention and deletion process.