Govern the workflow, not just the API call
Runtime governance for AI-agent workflows that can take consequential actions.
Arclasp evaluates governed actions against accumulated workflow state, pauses decisions that need a human, and preserves verifiable governance evidence.
Backend-authoritative. Fail-closed.
pip install arclaspA safe action can be part of an unsafe workflow.
An individual action can be acceptable on its own while the accumulated workflow crosses a business boundary. Arclasp evaluates the running Chain state—not only the latest governed action.
# Example with a configured $10,000 approval boundary
$4,000 -> allow
+$3,000 -> allow
+$4,000 -> require approval
Workflow exposure: $11,000Workflow-state governance catches what isolated checks miss.
Wrap consequential parts of a workflow in an Arclasp Chain. Governed actions are evaluated against the Chain state already accumulated, so policy can account for what happened earlier instead of treating every action as a fresh request.
A configured cumulative financial boundary can require human approval or deny the triggering action, depending on policy configuration.
import arclaspimport osarclasp.init(api_key=os.environ["ARCLASP_API_KEY"],backend_url="https://api.arclasp.com",)chain = arclasp.Chain("vendor-purchase")chain.add_financial_threshold(usd=10_000)async with chain:await chain.record_agent_action(agent_name="purchasing-agent",action_type="tool_call",action_name="record_vendor_commitment",payload={"vendor": "Northstar Components", "amount_usd": 4000},)# Customer application performs the real commitment only after# Arclasp permits the governed action.
What's in the box
The core controls for governing consequential AI workflows at runtime.
Workflow-state governance
Track Chain state and cumulative financial exposure across governed actions so policy can account for what happened earlier in the workflow.
Human approval gates
When configured policy requires review, governed execution pauses while an authorized human approves or denies the decision through supported approval flows.
Verifiable Chain Records
Valid completed Chains can produce signed, hash-linked, tamper-evident governance evidence with supported authenticated and scoped public verification paths.
Framework integrations
Use the framework-agnostic Python API or supported integrations for LangChain, LangGraph, CrewAI and MCP.
Keep your agent stack. Add governance around it.
Use the framework-agnostic Chain API or supported integrations for LangChain, LangGraph, CrewAI and MCP.
LangGraph
pip install "arclasp[langgraph]"from arclasp.langgraph import govern, governed_node
governed = govern(compiled_graph)LangChain
pip install "arclasp[langchain]"from arclasp.langchain import govern
governed = govern(executor)CrewAI
pip install "arclasp[crewai]"from arclasp.crewai import govern
governed = govern(crew)MCP
pip install "arclasp[mcp]"from arclasp.mcp import ArclaspMcpAdapter
adapter = ArclaspMcpAdapter(chain=chain, agent_name="my-tools")
await adapter.handle_tool_call(name, arguments, handler)Public beta. We tell you what's done and what isn't.
Live today: Python SDK on PyPI, open-source SDK repository on GitHub, hosted Arclasp governance backend, LangChain/LangGraph/CrewAI/MCP integrations, human approval workflows, Chain Records and verification.
Not in the current beta: mature self-hosting, non-Python SDKs, enterprise SSO/custom RBAC and universal result capture.
Read the full limitations page on docs.arclasp.com/limitations — we list everything that isn't built yet so you don't get surprised in production.
Open-source Python SDK
The Arclasp Python SDK is Apache-2.0. Inspect the client integration, framework adapters, verification helpers and failure behavior directly on GitHub.
The hosted governance backend and dashboard are not open source in the current beta.