examples

DashClaw Examples

Terminal-first governance examples showing the full decision loop: guard check, action recording, approval gate, and outcome tracking.

Two-terminal demo

The recommended way to run any example with an approval gate:

# Terminal 1: Run the agent
cd examples/openai-deploy-pipeline
npm install
node index.js

# Terminal 2: Approve when the gate fires
dashclaw approve act_<id shown in Terminal 1>

Examples

Example	SDK	Language	Governance Scenario
openai-governed-agent	OpenAI	Node.js	Customer refund email governance
anthropic-governed-agent	Anthropic	Node.js	Deployment agent held for human approval
claude-code-review-agent	Anthropic	Node.js	Security fix approval gate
codex-review-agent	Codex CLI	AGENTS.md	Same security-fix gate, governed via the Codex harness
openai-deploy-pipeline	OpenAI	Node.js	Production deploy approval with CLI
openai-agents-governed	OpenAI Agents SDK	Node.js	PII cleanup agent held for human approval
governed-chat-harness	Anthropic Messages API	Node.js	Chat runtime routing every tool call through guard
python-research-agent	None (simulated)	Python	File write governance
langgraph-governed	LangGraph	Python	StateGraph governance node pattern
crewai-governed	CrewAI	Python	@tool decorator governance pattern
autogen-governed	AutoGen	Python	Governed tool calls via the 4-step loop
pydantic-ai-governed	Pydantic AI	Python	Governed agent tool via the 4-step loop
vercel-ai-governed	Vercel AI SDK	Node.js	governed() wrapper for tool execute functions
managed-agent-governed	Anthropic (Managed Agent)	Python	Cloud-hosted agent with custom governance tools
managed-agent-mcp	Anthropic (Managed Agent)	Python	Same, via the MCP server — zero custom tools
kimi_dashclaw_test.py	Moonshot (OpenAI-compat)	Python	Governing a non-Anthropic model

openai-governed-agent

The original starter example. An OpenAI agent deploys a service to production. Shows guard, action, assumption, and outcome recording.

claude-code-review-agent

A Claude-powered agent reviews sample-auth.js for security issues. The file write triggers require_approval because the target path matches the auth risk pattern. Works without an Anthropic API key (uses simulated review output).

openai-deploy-pipeline

A CI/CD pipeline agent that runs pre-flight checks, gets an AI readiness assessment, and attempts a production deploy. The deploy action has risk 85 and is irreversible, which triggers the approval gate. Includes simulated rolling pod updates after approval.

python-research-agent

A Python agent that researches a topic and writes a report. Demonstrates the Python SDK governance flow. Requires no AI API key at all.

langgraph-governed

A LangGraph StateGraph with a governance_node that runs guard checks and records actions before the research node executes. Shows how to wire DashClaw into LangGraph's node-based execution model. Requires Python 3.10+ and the DashClaw Python SDK. No OPENAI_API_KEY needed.

crewai-governed

A CrewAI agent using the @tool decorator to wrap governance calls around tool execution. Demonstrates guard → create_action → update_outcome flow within CrewAI's tool abstraction. Requires Python 3.10+ and the DashClaw Python SDK. No OPENAI_API_KEY needed.

pydantic-ai-governed

A governed database-migration tool registered on a Pydantic AI agent via tools=[...]. Demonstrates the full 4-step loop including wait_for_approval HITL, plus the TestModel pattern for exercising the agent loop in tests. Requires Python 3.10+ and the DashClaw Python SDK. No LLM API key needed.

vercel-ai-governed

A generic governed() higher-order function that wraps any Vercel AI SDK tool execute in the 4-step loop (guard → createAction → waitForApproval → updateOutcome), applied to a support agent's lookup and refund tools. Node.js 20+ and the DashClaw Node SDK. No LLM API key needed.

Claude Managed Agents (MCP) ⭐ Recommended

managed-agent-mcp/ — The simplest way to govern a Claude Managed Agent. Uses DashClaw's MCP server — one config line gives the agent 17 governance tools and 3 resources. ~120 lines. Optionally pair with the dashclaw-governance skill (public/downloads/dashclaw-governance/) to teach the agent the governance protocol and load org-specific policies/capabilities automatically.

managed-agent-governed

A Claude Managed Agent running in Anthropic's cloud infrastructure with DashClaw as the governance layer. The agent has full access to bash, file I/O, and web search, but all external API calls, risky modifications, and significant decisions go through DashClaw custom tools (dashclaw_guard, dashclaw_invoke, dashclaw_record). Requires an Anthropic API key and a running DashClaw instance.

Governing non-Anthropic models

kimi_dashclaw_test.py shows DashClaw governing a Moonshot AI (Kimi) agent that uses an OpenAI-compatible endpoint. The governance loop — guard check, action record, assumption registration, outcome update — is identical regardless of which model drives the agent. Set MOONSHOT_API_KEY, DASHCLAW_BASE_URL, and DASHCLAW_API_KEY, then run python examples/kimi_dashclaw_test.py.

Market Intelligence Briefing (Full-Stack Demo)

Not an example you run externally — this seeds demo data directly into your DashClaw instance. Run node scripts/seed-demo-capabilities.mjs to create a knowledge collection, 5 capabilities, and 3 policies. See DEMO.md.

Prerequisites

All examples need:

• A running DashClaw instance (npm run dev from the repo root)
• DASHCLAW_API_KEY from your instance

Node examples additionally need Node.js 20+. Python examples need Python 3.10+.

Each example includes a .env.example file. Copy it to .env and fill in your keys before running.