You are under CARE.

Obligations

1. Use the correct Bearer for the intended record: Tool key for substrate register/bootstrap_rotate; Agent key for complete registration on one agent. A wrong Agent key configures a different agent—stop on identity mismatch.
2. Complete the full job sequence for your path. Do not stop after name, description, or other basic identity fields.
3. Never invent CARE origins, API keys, teams, or Accept/Discard. Ask the human when credentials or choices are missing. Humans Accept staged definition and Clearance merges; you do not.

Prerequisites

• CARE origin (example MCP URL: https://ai-managementsystem.com/mcp).
• A Tool key, an Agent key, or Path 0 if you can read the owner’s email and they have no CARE credentials yet.
• Prefer MCP tools at https://ai-managementsystem.com/mcp. Raw HTTP OpenAPI (https://ai-managementsystem.com/api/v1/docs, https://ai-managementsystem.com/api/v1/openapi.json) is optional for definition/assessments; SOP writes require MCP.

Which path? No key + email access → Path 0 then A then B. Otherwise call MCP initialize with your Bearer; read identity_type: tool → Path A, then Path B; agent → Path B only.

Path 0 — Owner bootstrap (email OTP → Tool key)

1. If you have no Tool key or Agent key, but you can read the owner’s email inbox, use Path 0. Otherwise ask the human for a Tool key (My tools) or Agent key.
2. Call POST https://ai-managementsystem.com/mcp with no Authorization Bearer. Call initialize; call tools/list — expect care_owner_bootstrap_start and care_owner_bootstrap_confirm.
3. Call care_owner_bootstrap_start with the owner’s email (optional agent_hint / platform_hint). Read the emailed one-time code (OTP). Confirming the OTP accepts CARE terms for agent-initiated signup.
4. Call care_owner_bootstrap_confirm with challenge_id and otp. Store the returned tool_key once in your vault. Do not invent passwords; the human sets a password later.
5. Switch Authorization Bearer to the Tool key. Continue with Path A, then Path B. When done, inform the human that CARE is ready (set password / Accept pending merges when needed). You do not Accept.

Path A — Tool key (substrate → register)

1. Confirm you have the CARE origin and a Tool key (from the human via Manage profile → My tools, or from Path 0). Do not invent credentials.
2. Call GET https://ai-managementsystem.com/mcp for discovery (no auth). Note under_care / under_care_md if present.
3. Attach POST https://ai-managementsystem.com/mcp with Authorization: Bearer <Tool key>. Call initialize; confirm identity_type is tool (stop if agent).
4. Call care_tool_whoami. Confirm platform, tool, tenant, and attributed user match expectations; stop on mismatch.
5. Call care_list_teams. Ask the human to choose a team by display name from that list—never invent teams; never use internal numeric ids. (Path 0 already ensured a default team exists.)
6. Call care_register_agent with name, external_correlation_id, chosen team (display name), and location (runtime control URL or e.g. “locally on machine”). On create: store agent_key once in your vault. On idempotent resume: identity only—no new secret; do not re-register hoping for a rotate.
7. On handoff failure after create only: call care_bootstrap_rotate_agent_key with agent_id (explicit). Never re-register expecting a new key.
8. Switch Authorization Bearer to the Agent key. Bootstrap clears on first successful Agent-key POST /mcp.
9. Immediately run Path B (complete registration) without waiting for human Accept of pending definition or Clearance merges.
10. Tell the human to Accept or Discard pending definition and Clearances in CARE when ready. You do not Accept.

Tool-key MCP tools (live tools/list under Tool key): care_tool_whoami, care_list_teams, care_register_agent, care_bootstrap_rotate_agent_key. They may not appear in https://ai-managementsystem.com/assets/mcp/tools.json (Agent-key catalog).

Path B — Agent key (complete registration)

1. Call MCP initialize. Confirm identity_type is agent. Read agentScope { agent_id, tenant_id } and instructions. Stop on identity mismatch before any write.
2. Complete CARE registration for this agent: do not stop after name, description, or other basic identity fields—registration is incomplete without (1) staged definition graphs, (2) mandatory Clearances, (3) improvement recommendations reconcile per required Clearance, (4) agent SWOT matrix + top recommendations, (5) SOP Markdown, and (6) process/SOP improvement recommendations.
3. When the human says self-register, complete registration, or equivalent: run the full sequence below.
4. If CARE host or this agent's API key is missing, ask the human for them; do not invent credentials.
5. Before any stage/merge/SOP/recommendations/SWOT/process-recommendation write: read initialize.agentScope and care_get_definition; confirm agent identity matches your target. If it does not match, stop immediately—do not write.
6. Stage definition: care_get_import_schema → care_stage_definition with vendor_data.care.role_model and vendor_data.care.process_model (items + evidence-based connections; unique (section, text)). Set agent.personal_agent true|false from evidence (true = Agent Owner alone fully accountable for manage/operate/control; false = distinct ownership roles). Lifecycle (agent.status): if the agent is set up to run repeatedly (scheduled/recurring), set active; otherwise set in_development—unless care_get_definition already shows active (never downgrade Active).
7. Clearances (always on onboarding / self-register): care_list_assessment_catalog → care_merge_assessment for every pack on this agent's plan (required and optional). Never skip a catalog pack; use JSON null for unknown items. Lifecycle Active still requires the five mandatory Clearances — Use case viability (agentic-usecase-viability), Process capability (agentic-process-capability), Risk classification (agentic-risk-classification), Instance maturity (agent-instance-capability), Security (agentic-security) when on plan.
8. Improvement recommendations (per catalog Clearance after merge): care_get_assessment_recommendations → judge previous as implemented or open → care_set_assessment_recommendations (optional status new; ≤5 open; importance/urgency 1–5; evidence-only, no padding, no delete). Write-through; do not invent filler. Failed or unknown Clearance points (false/null) are assessment evidence only—not one recommendation per point, and not automatic Critical/Do-first ToDos. Recommend distinct strategic control/capability themes for owner or executive attention; do not confuse operational bugs, heartbeats, checklists, or runtime triage with CEO-level actions (put how-to fixes in SOP/process recommendations). Deduplicate: one open per distinct theme—if the same remediation already exists (same or near-same title/intent) on this Clearance or another Clearance for this agent, do not add another new open; keep/strengthen the canonical open, mark reworded twins implemented, or omit. Re-run with Clearances after a significant definition change.
9. Agent SWOT matrix (after Clearance recommendations): care_get_agent_swot_matrix → judge previous SWOT findings and top recommendations as implemented or open from all Clearance open points / outcomes → care_set_agent_swot_matrix (swot + recommendations; SWOT type S|W|O|T; ≤3 open per type; ≤3 open top recommendations; importance/urgency 1–5; evidence-only, no padding, no delete). Deduplicate reworded twins within SWOT and top recommendations. Write-through.
10. After a significant definition change (jobs, permissions, process, off-limits, or authority), re-run every catalog Clearance, their recommendation reconcile, and the agent SWOT matrix so they stay current with the definition—do not leave stale clearance answers.
11. SOP: care_get_sop → follow authoring_instructions and template → care_update_sop (full markdown= or section=+content= PATCH; exact required_headings only).
12. Process improvement recommendations (after SOP; the agent’s process is its SOP): care_get_sop (and care_get_definition if needed) → care_get_process_recommendations → judge previous as implemented or open → care_set_process_recommendations (optional status new; ≤5 open; importance/urgency 1–5). Cover operating HOW (steps, checkpoints, handoffs) and the WHAT of the work for agentic control (reliability, low ambiguity, double binds, conflicting instructions, unclear outcomes)—not Clearance remediations and not a substitute for care_update_sop. Evidence-only; no padding; no delete; deduplicate reworded twins. Write-through.
13. Tell the human to Accept pending definition and assessment merges in CARE. You do not Accept.

Follow CARE MCP initialize.instructions for graph semantics, evidence discipline, and assessment wire encoding.

Endpoints

Done checklist

• Correct Agent key bound to the intended agent_id (initialize.agentScope matches care_get_definition).
• Definition staged with graphs under vendor_data.care; location set; not only name/description.
• Every Clearance pack on the plan merged (required and optional).
• Clearance improvement recommendations, agent SWOT, SOP, and process improvement recommendations written per Path B.
• Human notified to Accept/Discard pending definition and Clearance merges.
• (Path 0) Owner bootstrapped via email OTP if no prior CARE credentials; Tool key stored once.
• (Path A) Agent created via Tool key; Agent key stored from create or bootstrap_rotate only.
• Human informed that CARE is ready (set password / Accept when needed).