Starknet Agentic
Production-grade Starknet agent infrastructure: secure Cairo contracts, MCP/A2A runtimes, x402 payment helpers, end-to-end examples, and installable agent skills.
This repository is for developers building agents that need Starknet-native wallets, policy-enforced execution, on-chain identity, and composable tool access.
Install & Use
Scaffold a Starknet agent
npx @starknetfoundation/create-starknet-agent@latest
The scaffolder detects supported agent environments and wires Starknet integration into a new project.
Install a skill
npx skills add keep-starknet-strange/starknet-agentic/skills/cairo-auditor
Codex public GitHub install:
CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
python3 "$CODEX_HOME/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
--repo keep-starknet-strange/starknet-agentic \
--path skills/cairo-auditor \
--ref main
Codex reproducible install:
CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
python3 "$CODEX_HOME/skills/.system/skill-installer/scripts/install-skill-from-github.py" \
--repo keep-starknet-strange/starknet-agentic \
--path skills/cairo-auditor \
--ref <commit-sha>
Claude Code marketplace install:
/plugin marketplace add keep-starknet-strange/starknet-agentic
/plugin install starknet-agentic-skills@starknet-agentic-skills --scope user
For Codex, Claude Code, and pinned install flows, use the deterministic skill quickstart:
Run the MCP server from source
From a source checkout after pnpm install:
pnpm --filter @starknetfoundation/starknet-agentic-mcp-server build
node packages/starknet-mcp-server/dist/index.js
Production deployments should use proxy signer mode rather than in-process private keys. See packages/starknet-mcp-server and docs/security/SIGNER_API_SPEC.md.
What Is Included
| Area | Path | Purpose |
|---|
| Cairo contracts | contracts/ | Agent accounts, ERC-8004 registries, session-account primitives, and registry experiments |
| TypeScript packages | packages/ | CLI scaffolder, MCP server, A2A adapter, agent passport helpers, onboarding utilities, prediction scanner, and x402 helpers |
| Skills | skills/ | Public agent skills for Cairo auditing, Starknet wallets, DeFi, identity, testing, deployment, optimization, and SDK usage |
| Examples | examples/ | Reference agent flows covering onboarding, identity, MCP loops, DeFi, carry monitoring, controller calls, and cross-chain demos |
| Datasets and evals | datasets/, evals/ | Cairo audit/evaluation corpora and deterministic benchmark material |
| Docs | docs/ | Architecture, roadmap, deployment status, security runbooks, and launch material |
| Website | website/ | Documentation site source |
Architecture
flowchart TB
Agent["Agent runtime<br/>(Codex, Claude Code, OpenClaw, custom app)"] --> Tools["Tool layer<br/>MCP, A2A, skills"]
Tools --> Signer["Signer boundary<br/>direct for local dev, proxy for production"]
Signer --> Starknet["Starknet"]
Starknet --> Accounts["Agent/session accounts<br/>policy enforcement"]
Starknet --> Registries["ERC-8004 registries<br/>identity, reputation, validation"]
Tools --> Packages["TypeScript helpers<br/>passport, onboarding, x402, scanners"]
The recommended launch profile is self-custodial and no-backend:
- users run the agent runtime locally or on their own infrastructure
- account contracts enforce transaction policy on-chain
- production signer custody lives behind an explicit proxy/KMS/HSM boundary
Core Components
Contracts
