Drop-in agent rule

GPT Actions

The OpenAPI path lets a GPT call CrawlDex as a read-side tool before it suggests a browser task or user handoff.

Install or paste

Action import

text

Import this OpenAPI URL into the GPT/action tool setup:

https://crawldex.com/openapi.json

Instruction for the model:
Before recommending or attempting a public website task, call resolveIntent when the task is vague, then call preflightTask with the target site and task. Send x-crawldex-channel: pack-gpt-actions when the action builder supports headers. Use agent_instruction as the final action posture. Do not call reporting endpoints unless the user explicitly asks to submit a sanitized public observation.

Privacy boundary

Keep CrawlDex public-only.

Do not submit private account data, payment data, messages, screenshots with sensitive content, cookies, tokens, raw traces, local files, or logged-in-only pages. Stop before irreversible work unless the user is present and explicitly authorizes it.

OpenAPI smoke test

bash

curl -sS -H "x-crawldex-channel: pack-gpt-actions" https://crawldex.com/openapi.json | jq '.paths["/api/v1/preflight"].post.operationId'

Report redacted outcome

bash

curl -sS -X POST https://crawldex.com/api/v1/runs \
  -H "content-type: application/json" \
  -H "x-crawldex-agent-key: $CRAWLDEX_AGENT_KEY" \
  -H "x-crawldex-channel: pack-gpt-actions" \
  -d '{"site":"netflix.com","task":"subscriptions.cancel","outcome":"success_with_handoff","friction":["login_required"],"source_tier":"anonymous_report","evidence":{"artifact_types":["redacted_agent_report"]}}'