IAGENT-AUTOPILOT

iAgent Autopilot is the execution and safety layer on top of the official Injective MCP Server — built for the Injective Solo AI Builder Sprint (May 2026).

Most AI agents on Injective read the chain or suggest trades. Autopilot executes within limits the operator defines: max notional, leverage, daily loss, allowed markets, and a one-click kill switch. Five specialized agents coordinate over a pub/sub pipeline (event → proposal → verdict → execution → audit), each persisted and streamed live to a Next.js workbench.

How AI is used: Groq Llama 3.3 70B (Watcher, Risk sanity check), Claude Sonnet 4.5 (Analyst, Auditor), Claude Haiku (strategy parse). Risk runs on a different model than the Analyst so the proposer never grades its own homework. The Executor has no LLM — it is the only component with MCP write access.

How Injective is integrated: All chain actions flow through MCP tools (trade_open, trade_close, transfer_send, balances, positions, etc.) on testnet. Public demo runs simulator + dry-run for judge safety; optional DEMO_REAL_TX produces an explorer-verifiable transfer_send for proof of execution.

Shipped for the sprint

• Five-agent runtime (FastAPI, SQLite, WebSocket) with deterministic risk gates, kill switch, and dry-run by default

• Operator dashboard on Vercel — strategy editor, agent fleet, decision pipeline, demo scenarios, live audit stream

• Public deploy — Render API + Vercel UI, API-key proxy, CORS, cold-start “waking engine” UX for judges

• Demo scenarios — Funding Reversion (full pipeline), Risk Block, Kill Switch; judge onboarding (pulse + hint)

• On-chain proof path — optional testnet transfer_send demo with explorer links

• Documentation — README (AI + Injective integration), DEPLOY.md, screenshots, 90s demo

• Tests — 60+ runtime unit tests; dashboard production build verified

Current limitations (honest)

• Render free tier cold-starts (~30s first load); dashboard retries automatically

• Public deploy does not run Node MCP on Render — judges use simulator + dry-run (by design)

• Live perp trading needs MCP + wallet on a VPS or local machine

• Risk Block scenario depends on Analyst + Groq behavior; deterministic limits always apply

Next steps (post-hackathon)

1. Mainnet hardening — stricter defaults, position caps, slippage checks

2. Extend MCP tool registry — same pipeline for staking, bridge, limit orders (28 tools today)

3. Strategy marketplace — import/export strategy JSON

4. Backtesting — replay historical events through Watcher → Analyst → Risk without execution

5. Multi-wallet — N strategies × N accounts with shared audit stream

6. Hosted MCP — run Injective MCP beside the API for live demo without local setup

   Value for the Injective ecosystem

   Autopilot is infrastructure, not a one-off bot: a reusable governed execution shell for any agent that calls Injective MCP — so builders get safety, audit, and UX without rebuilding coordination from scratch.

   ---

   It trades. You stay in control. — iAgent Autopilot: five AI agents, hard limits, Injective MCP, full audit trail.

NOT YET BUT VERY SOON