ClawMind

What ClawMind does

ClawMind runs multi-agent due diligence on Web3 projects and anchors every report on 0G Chain, so audit results can be verified later instead of silently rewritten.

A user submits a project description. Eight specialized agents analyze it end-to-end. An adversarial Critic challenges the other agents' assumptions. The final report is hashed, signed by an authorized operator via EIP-712, and recorded in the AnalysisRegistry smart contract on 0G Chain mainnet.

The 8-agent pipeline

1. Memory Retrieval - semantic search over prior analyses stored through the 0G-backed memory index

2. Planner - decomposes the task into research, risk, architecture, and verification tracks

3. Researcher - extracts facts, assumptions, and external risk signals

4. Risk Agent - identifies custody, governance, operational, and economic risks

5. Architect - proposes mitigations and architecture-level improvements

6. Critic - adversarially challenges prior conclusions and flags unresolved issues

7. Final Synthesis - reconciles all outputs into a score, recommendation, and report

8. Memory Writer - appends the completed run to persistent semantic memory

Inference is routed through 0G Compute. Memory retrieval uses 384-dimensional embeddings with all-MiniLM-L6-v2.

What makes it different

- Multi-agent reasoning pipeline. ClawMind separates planning, research, risk assessment, architecture review, adversarial critique, final synthesis, and memory writing into distinct pipeline stages.

- Adversarial Critic with measurable impact. Each unresolved Critic challenge mathematically lowers the score: high -15, medium -7, low -3. The math is visible in the UI and reproducible.

- EIP-712 signed receipts. Reports are signed by an authorized operator before being written on-chain. The smart contract verifies the signature on every write and rejects unauthorized submissions.

- Persistent semantic memory. Each analysis is stored as a verifiable snapshot. Future runs retrieve relevant precedents through cosine similarity over 384-dimensional embeddings.

- MCP integration. The full pipeline is exposed as a Model Context Protocol server. Any MCP-compatible client, including Claude Desktop and Cursor, can invoke ClawMind with two tools: analyze_web3_project and get_recent_analyses.

Track 1 alignment

ClawMind covers the core requirements of Track 1: Agentic Infrastructure & OpenClaw Lab.

- OpenClaw orchestration - openclaw.yaml defines the full 8-step pipeline, skills, artifacts, and security policies. Served at /api/openclaw/manifest.

- 0G Compute - agent inference is routed through 0G Compute.

- 0G Storage - decision reports and memory records are persisted with verifiable 0g:// URIs.

- 0G Chain - AnalysisRegistry.sol records each analysis with root hash, score, recommendation, storage URI, and operator signature.

- Long-context memory — runtime-grown semantic memory with active retrieval and public stats.

Audience

ClawMind is built for DAO contributors evaluating external proposals, Web3 VCs needing reproducible due diligence, and protocol teams looking for an independent pre-screen with a verifiable audit trail.

What it isn't

ClawMind is a due-diligence aid, not a formal audit or exploit detector. It does not execute transactions, manage user funds, or replace human security review. On-chain integrity proves that a report hash was signed and recorded; it does not prove the report is correct.

What was built during the hackathon

Smart contract - AnalysisRegistry.sol

Deployed twice on 0G mainnet. The first iteration 0x01c9...4D2) was open-write; the second 0x08a9c275...e2b1) adds EIP-712 operator authentication. The contract verifies a typed-data signature on every recordAnalysis() call and rejects writes from unauthorized signers. Operator 0x9A0C8040...99F8 is currently authorized. Source: contracts/AnalysisRegistry.sol.

8-agent pipeline

Built from scratch: Memory Retrieval → Planner → Researcher → Risk → Architect → Critic → Final Synthesis → Memory Writer. Each agent has its own skill, prompt, and model assignment. Final Agent applies explicit score anchors plus Critic severity penalties (high −15, medium −7, low −3), so calibration is measurable, not vibes.

0G Compute integration

Agent inference is routed through 0G Compute. The production deployment currently uses deepseek/deepseek-chat-v3-0324 as the primary model route, with model routing/fallback logic in the codebase.

Persistent semantic memory

Memory index lives on 0G Storage with versioned, immutable snapshots. Each analysis appends a record. Retrieval uses cosine similarity over all-MiniLM-L6-v2 embeddings (384 dimensions).

MCP server

Deployed as a separate Vercel project apps/mcp-server). Two tools: analyze_web3_project and get_recent_analyses. Authentication via X-MCP-Client-Id header with 60-second rate limit per client. The MCP server is a thin wrapper that calls the existing /api/analyze endpoint, so MCP-initiated analyses get the same EIP-712 signing and on-chain recording as web-initiated ones.

Production UX

Core screens: landing, /analysis or /analyze, live pipeline/report view, /stats, and /judge. Live pipeline shows real-time agent output, model strings, on-chain status (task hash → root hash → signature → tx), and a Critic moment where unresolved challenges visibly lower the score.

Verifiable end-to-end loop

A user can submit a task, see it analyzed, retrieve the report by storage URI, and verify the on-chain hash matches the storage content. The "Retrieve Report" widget on the report page accepts any 0g:// URI or root hash and reconstructs the persisted report.

Numbers so far

- More than 35 analyses recorded on 0G mainnet

- Current signed registry writes use EIP-712 operator authentication

- More than 18 memory records, including 15 runtime-grown records

- Score distribution covers INVESTIGATE_MORE and NO_GO in recent runs, with GO present in historical calibration runs

- Critic raised 54 challenges across 14 tracked runs, average score adjustment about -26

Honest limitations

- The "OPEN SOURCE" claim on the landing page is backed by a public repo, but the LICENSE file is permissive (MIT) and minimal.

- Score floor is currently 0; a few analyses hit exactly 0, which reads as a fallback. In a follow-up we will clamp at 5+ for non-trivial input.

- Memory growth is real but small. The system needs more runs before semantic retrieval produces high-similarity matches consistently.

Not raised. ClawMind is a hackathon project built for Track 1 of the 0G Hackathon, currently in closed beta. We're focused on validating the multi-agent due-diligence pipeline with real Web3 reviewers (DAO contributors, protocol teams, Web3 VCs) before any fundraising conversation.

Open to grants and ecosystem support from 0G or aligned partners. Not actively pitching to VCs at this stage.