webhooks:

- name: "slack-alerts"

url: "https://hooks.slack.com/services/YOUR/WEBHOOK/URL"

events: [deny, escalation, risk_threshold]

format: slack

- name: "discord-log"

url: "https://discord.com/api/webhooks/YOUR/WEBHOOK"

events: [deny, escalation, session_start, session_end]

format: discord

- name: "custom-siem"

url: "https://your-siem.internal/ingest"

events: [all]

format: json

headers:

Authorization: "Bearer ${SIEM_TOKEN}"

Event Types

You can subscribe to any combination of SafeClaw events:

• deny — An action was denied by policy. Useful for security monitoring.
• escalation — An action was escalated for human review. Critical for teams that share approval responsibilities.
• risk_threshold — A session's risk score crossed a warning or critical threshold.
• budget_warning — A budget pool reached its warning threshold.
• budget_exhausted — A budget pool was fully consumed.
• session_start / session_end — Agent sessions beginning and ending, with summary statistics.
• all — Every event. Best for SIEM ingestion where you want the complete firehose.

Payload Customization

The default payloads include the event type, timestamp, session ID, action details, and the decision reason. You can customize payloads using a template system:

webhooks:

- name: "minimal-slack"

url: "https://hooks.slack.com/..."

events: [deny]

template: |

:warning: Agent denied: {{action.type}} on {{action.target}}

Reason: {{decision.rule}}

Templates support Handlebars-style interpolation with access to the full event data. This is particularly useful for custom endpoints that expect specific payload formats.

Reliability

Webhook delivery is best-effort with retries. If a webhook endpoint returns a non-2xx status, SafeClaw retries with exponential backoff (1s, 5s, 30s, 5m) up to 5 times. Failed deliveries after all retries are logged locally and reported in the dashboard.

When SafeClaw is offline, webhook payloads are queued to disk and delivered when connectivity returns. The queue is bounded to prevent unbounded disk growth — oldest events are dropped first when the queue reaches capacity.

Security