Open-Source AI Agent Framework with Cryptographic Verification Claude Agent SDK + Pluggable Skills + Bitcoin-anchored integrity.
OpenClaw is a monorepo of composable building blocks for teams that want to:
- Ship an AI agent that invokes a Claude model with domain-specific Skills
- Produce cryptographically verifiable output (SHA-256 + Merkle + OpenTimestamps)
- Operate locally, on a server, or at the edge — same primitives everywhere
It is not a hosted service, not a fund, not an investment platform. It is a software library you run yourself.
# 1. Install
npm install @openclaw/agent @openclaw/skills @openclaw/verify
# 2. Scaffold a project
npx openclaw init my-agent
cd my-agent
# 3. Plug in a skill (write markdown + frontmatter)
cat > skills/hello.skill.md <<'EOF'
---
name: hello
description: Greet the user by name
---
When asked to greet someone, reply with a short haiku
that includes their name.
EOF
# 4. Run
export ANTHROPIC_API_KEY=sk-ant-...
npx openclaw run --skill hello "greet Neo"The agent will load the skill, call Claude, and return the response. Every step is hashed; running npx openclaw verify will re-check the full trace against a Merkle root.
| Package | Purpose | Status |
|---|---|---|
@openclaw/verify |
SHA-256 canonical JSON · Merkle tree · OpenTimestamps proof · chain utilities | ✅ v0.1 |
@openclaw/agent |
Thin wrapper around Claude Agent SDK · plugs in Skills · emits verifiable traces | 🚧 v0.1 scaffold |
@openclaw/skills |
Markdown + frontmatter Skill loader · registry · sandboxed tool interface | 🚧 v0.1 scaffold |
@openclaw/parser-cme |
Parse Chinese-broker CME futures settlement statements into canonical JSON | ✅ v0.1 (reference data only) |
data/reference-datasets/ ships real-world datasets that have been processed through the verify pipeline, so the technical claims can be independently checked:
| Dataset | Scope | Purpose |
|---|---|---|
q1-2026/ |
Q1 2026 · 6 independent proprietary futures accounts · CME NQ E-mini · 1,247 fills · 63 trading days | Stress-test the SHA-256 / Merkle / OTS pipeline on institutional-scale settlement data |
These datasets are technical fixtures, not performance claims. Account owners are anonymised (Partner Account A-F). The point is: these hashes reproduce deterministically from the raw data we publish, and the Merkle roots are anchored on the Bitcoin blockchain. Anyone can run npx @openclaw/verify check q1-2026 to confirm.
No dataset on this site constitutes an investment track record, a recommendation, or an offer. OpenClaw does not manage money, does not solicit investment, and is not a licensed financial entity.
Three commitments:
-
Verifiable by construction. Every output has a canonical JSON form and a hash. Every hash aggregates into a Merkle root. Every root gets Bitcoin-anchored timestamps. You don't have to trust us — you can re-derive everything from raw inputs.
-
Skills as first-class plugins. Your domain knowledge lives in
.skill.mdfiles — not buried in prompts or tangled into code. Swap, version, and share skills like npm packages. -
No lock-in. Apache 2.0. No telemetry. No server dependency. Fork and self-host.
openclaw/
├── packages/
│ ├── verify/ @openclaw/verify · core crypto primitives
│ ├── agent/ @openclaw/agent · Claude Agent SDK wrapper
│ ├── skills/ @openclaw/skills · skill loader + registry
│ └── parser-cme/ @openclaw/parser-cme · CME settlement parser
├── apps/
│ └── web/ docs + reference-data browser (Next.js 14)
├── data/
│ └── reference-datasets/
│ └── q1-2026/ 6 files · canonical JSON · Merkle roots · OTS proofs
├── scripts/
│ ├── verify-cli.mjs `npx openclaw verify` entry point
│ └── gen-reference-data.mjs
└── docs/ long-form design notes + RFCs
- Issues & PRs: https://github.com/gmgai-lab/openclaw
- Discord: (link once community channel is live)
- Apache 2.0 license — commercial use permitted, attribution appreciated
OpenClaw was extracted from GMG·AI, a research project exploring AI-assisted quantitative workflows. The verification primitives and agent infrastructure proved to be the most reusable layer, so we split them into this standalone open-source framework.
GMG·AI's public site (gmgai.org) now presents OpenClaw as its primary deliverable. The project does not accept investment, does not manage client funds, and is not a licensed financial entity in any jurisdiction.