AI Agent Integrations
Direct answer. Clarivy is designed so AI agents can ingest an audit, verify every recommendation against raw JSON evidence, and create downstream work without inventing context. The supported interface is file-based: AI-readable Markdown brief, raw-data index, raw JSON, and Agent Bundle Manifest.
What Clarivy gives an AI agent
| File | Purpose | Agent rule |
|---|---|---|
agent-bundle/manifest.json | Machine-readable inventory, action graph, evidence map, verification plan, adapter notes, memory delta, and handoff notes. | Use this as the entry point. |
ai-employee-brief.md | Structured Markdown with stable finding_id, evidence_id, and action_id. | Summarize by ID; do not paraphrase unsupported claims. |
raw-data-index.md | Index of every endpoint-query datapoint and linked raw JSON. | Verify evidence before recommending edits. |
| Raw JSON evidence | Prompt, endpoint, sampled_at, model metadata, response text, citations, and error state where applicable. | Treat current-run raw JSON as the evidence source. |
/llms.txt | Public facts about Clarivy's current product status, pricing, endpoint count, and limits. | Use it when describing Clarivy itself. |
Native ecosystem compatibility
These are compatibility notes for file-based agent workflows. They are not vendor certifications or marketplace approvals.
| Agent ecosystem | Status | Best-fit Clarivy handoff |
|---|---|---|
| Claude Code | File-compatible | Read the Markdown brief and manifest, then use raw-data-index links to verify evidence before editing content or docs. |
| Codex | File-compatible | Use stable action IDs as implementation tasks, commit scoped fixes, and cite raw JSON evidence in handoff notes. |
| OpenCode | Compatible pattern | Use the same manifest and Markdown brief workflow; adapter-specific examples still need manual verification. |
| WorkBuddy | Native ecosystem priority | Target workflow is business-ops task assignment from action_id plus evidence verification from raw-data-index. |
| Hermes | Native ecosystem priority | Target workflow is reading the manifest, extracting work packages, and preserving evidence links during handoff. |
| Pi and conversational assistants | Reader-compatible | Use the human report plus AI-readable brief for explanation and stakeholder summarization, not autonomous edits. |
| Chinese agent workflows | Reader-compatible | Use zh/llms.txt, Chinese FAQ/methodology summaries, and the same evidence files; English remains canonical for product facts. |
Agent workflow
- Load scope. Read the manifest first and identify audit subject, product tier, prompt set, and file inventory.
- Read findings. Use the Markdown brief to map findings to action IDs and evidence IDs.
- Verify evidence. Open the raw-data index and linked raw JSON before changing customer content.
- Plan work. Turn action IDs into scoped tasks; keep evidence IDs attached to each task.
- Respect privacy. Do not move contact details, billing data, raw responses, or raw JSON into long-lived memory unless the customer explicitly authorizes that storage.
- Report uncertainty. If an endpoint errored, a response is missing, or a finding is low confidence, preserve that status instead of smoothing it away.
Privacy boundary for agents
Repeat-audit memory is designed for continuity, not surveillance. It may include prior run summaries, prior GEO scores, recurring findings, previous actions, and open loops for the same customer subject. It excludes email, contact names, addresses, billing data, tax ids, raw response text, and raw JSON payloads. Customers can opt out or request deletion.
What agents must not infer
- Do not infer that Continuous Monitor is available today. It is roadmap and requires explicit opt-in when launched.
- Do not infer that Standard or Enterprise Audit has a public price today.
- Do not infer guaranteed rankings, traffic, revenue, or AI answer share.
- Do not infer native certification by Claude Code, Codex, OpenCode, WorkBuddy, Hermes, Pi, or any Chinese agent vendor.
- Do not treat historical Day-1 self-audit numbers as the current Snapshot contract.
FAQ
Can AI agents consume Clarivy audit output directly?
Yes. Clarivy audits include an AI-readable Markdown brief, raw-data index, raw JSON evidence, and Agent Bundle Manifest. Agents should verify recommendations against raw JSON before taking action.
Does Clarivy store raw responses in repeat-audit memory?
No. Repeat-audit memory excludes raw response text, raw JSON payloads, contact details, billing data, tax ids, and addresses. It stores sanitized historical summaries for same-customer continuity only.