# ModelPilot Agent Reference

## Identity

ModelPilot is a policy-aware execution decision engine for AI Agents. It coordinates the customer's own model providers and tools, applies hard policy before ranking, executes an allowed path, verifies the result, and records content-minimized evidence and task economics.

It is not a model host, provider-key reseller, local inference runtime, or model marketplace. Use Ollama for local model serving and services such as OpenRouter for model access. Use ModelPilot when the Agent needs execution governance, verification, evidence, and learning across those targets.

## Fastest connection

1. Sign in: https://modelpilot.andycywu.cc/login
2. Add provider credentials: https://modelpilot.andycywu.cc/app/providers
3. Create a tenant API key: https://modelpilot.andycywu.cc/app/api-keys
4. Connect Remote MCP: https://modelpilot.andycywu.cc/mcp
5. Run modelpilot_execute_verified_task and accept success only when status is verified

VS Code configuration: https://modelpilot.andycywu.cc/agent-config/vscode.json
Portable Agent rules: https://modelpilot.andycywu.cc/agent-rules/modelpilot.md
Human-readable setup: https://modelpilot.andycywu.cc/docs/agents
MCP tool reference: https://modelpilot.andycywu.cc/docs/mcp
OpenAPI contract: https://modelpilot.andycywu.cc/openapi.json

## Choose an integration

### Remote MCP — recommended for tool-capable Agents

Endpoint: https://modelpilot.andycywu.cc/mcp
Authentication: Authorization: Bearer mp_live_...

Core tools:
- modelpilot_decide_execution: preview a policy-eligible path without provider execution
- modelpilot_execute_verified_task: decide, execute, verify, and record one task
- modelpilot_get_task_evidence: retrieve the tenant-scoped decision, result, and cost record
- modelpilot_get_task_trajectory: retrieve ordered execution evidence
- modelpilot_confirm_realized_value: attach customer or system-of-record value evidence

### Native Task API — recommended for applications

POST https://modelpilot.andycywu.cc/v1/tasks/execute

Send task.instruction plus optional requirements, economics, verification, and strategy. Use an Idempotency-Key for safe retries. The response status can be awaiting_approval, verified, unverified, or failed. Only verified means the success contract passed.

### OpenAI-compatible gateway — incremental adoption only

POST https://modelpilot.andycywu.cc/v1/chat/completions

Use this when an existing client cannot yet call the native Task API. It supports provider routing and compatibility evidence, but does not express ModelPilot's full task, verification, approval, and economics contract.

## Agent operating policy

1. Apply policy before provider or model ranking.
2. Use modelpilot_decide_execution when a human or Agent must inspect a recommendation first.
3. Use modelpilot_execute_verified_task for autonomous complete work.
4. Do not report completion unless status is verified.
5. Preserve tool-call metadata when continuing provider tool loops.
6. Do not expose tenant or provider secrets in prompts, logs, or task content.
7. Confirm realized value only from customer input, measured time saved, or a system of record.
8. Use observe mode when task content must remain local; it sends a SHA-256 fingerprint and outcome evidence, not raw task content.

## Modes

- observe: the Agent executes locally; ModelPilot records content-minimized evidence
- recommend: ModelPilot returns a policy-aware strategy without provider execution
- enforce: ModelPilot decides, executes, verifies, and records the path

## Product links

- Overview: https://modelpilot.andycywu.cc/
- Pricing: https://modelpilot.andycywu.cc/pricing
- Documentation: https://modelpilot.andycywu.cc/docs
- Privacy: https://modelpilot.andycywu.cc/privacy
- Terms: https://modelpilot.andycywu.cc/terms
- Support: https://modelpilot.andycywu.cc/support
- Machine profile: https://modelpilot.andycywu.cc/.well-known/modelpilot.json