Silo

Demo

About

The problem it solves SILO: Secure AI-Agent Workflows on 0G SILO addresses a core challenge in AI-agent development on 0G: teams can build agent workflows quickly, but struggle to keep memory private, behavior auditable, and automation safe. Key Use Cases Encrypted Memory: Agents store sensitive data encrypted by default before writing to 0G. Session Attestation: Prove agent actions with attestation metadata and Merkle-based evidence. MCP Integration: Connect agent capabilities directly to Claude or Cursor without custom glue code. DeFAI Safety Flows: AI produces structured plans with guardrails, risk previews, and explicit user approval. Multi-Agent Coordination: Shared memory, heartbeat monitoring, and autonomy levels (monitor / suggest / autonomous) with controlled execution modes. How SILO Makes Work Easier and Safer Faster Onboarding: Scaffold CLI, doctor checks, and copy-paste setup docs get teams running quickly. Safer Defaults: Explicit approval gates, dry-run DeFAI mode, and no automatic broadcast by default. Better Debugging and Trust: Live event logs, auditable artifacts, and reproducible examples help debug and verify workflows. Use of AI tools and agents SILO: AI Agents on Secure 0G Infrastructure SILO treats AI agents as first-class operators, running over secure, 0G-backed infrastructure. How AI is Used Through MCP, agents (Claude/Cursor) call SILO tools directly to: Encrypted memory operations: vault_store, vault_retrieve, session_commit DeFAI planning and safety control: defai_plan, defai_approve Shared-memory coordination and autonomy status How Agents Work Together SILO supports multi-agent collaboration via shared memory channels and routing primitives. Sub-agents can be created for specialized tasks, coordinated through: Channel-based memory exchange Heartbeat-driven system health & maintenance Autonomy levels (monitor, suggest, autonomous) depending on desired control Safety & Control Model AI outputs are structured and constrained by guardrails (slippage, timeout, allowlist, notional caps) Explicit user approval is required before execution handoff DeFAI mode is dry-run only — no automatic transaction broadcast in this MVP Auditability Memory and decision artifacts are encrypted and persisted on 0G Session activity is attested with Merkle-based proofs, making agent behavior verifiable without exposing plaintext data Summary AI agents generate and operate workflows, while SILO provides the secure memory, coordination, and verification layer that makes those workflows safe, auditable, and reproducible on 0G.

How it's made

Challenges we ran into Real Implementation Hurdles & Fixes in SILO 1️⃣ 0G SDK / Chain Compatibility Drift Issue: SDK and chain versions were drifting, causing inconsistent behavior. Fix: Added a compatibility layer in the storage client to normalize submission payloads, keeping uploads reliable across the current testnet behavior. 2️⃣ MCP JSON Parse Failures in Claude/Desktop Issue: MCP clients require clean protocol frames on stdout. Normal logs contaminated output, causing …is not valid JSON errors. Fix: Redirected non-protocol logs to stderr in the MCP runtime. Tightened setup docs to prevent misconfigured launches. 3️⃣ Version Drift Between Local Repo and NPM Latest Issue: New MCP tools (defai_plan, defai_approve) existed locally but not in older published packages, making them appear “missing.” Fix: Published updated package versions. Updated config guidance to explicitly pull the latest package in MCP arguments. 4️⃣ Session Commit Instability on Testnet Issue: session_commit could revert due to testnet/RPC volatility and process nonce contention. Fix: Enforced safer run patterns (single writer process, manual-start guard). Added clear fallback behavior in demo flows. Documented troubleshooting paths for reproducibility. ✅ Outcome: These fixes significantly improved real-world reliability and strengthened the developer experience, making SILO safer, more stable, and easier for other teams to adopt.