Skip to main content

​
Arclasp

Your agents act together. Govern them that way. Arclasp is runtime governance for AI-agent workflows. A workflow becomes an Arclasp Chain. Your application records the consequential actions it wants governed, and the Arclasp backend evaluates each governed action against the workflow state accumulated so far. The backend can return four policy outcomes:
  • allow
  • allow_with_flag
  • require_approval
  • deny
When approval is required, the workflow waits for an authorized human decision. When a valid Chain completes, Arclasp can produce signed, hash-linked evidence that can be verified later. The customer application still executes its own tools. Arclasp is the governance decision boundary around those instrumented actions; it is not the tool executor.

​
Why workflow state matters

A single action can look acceptable while the workflow becomes risky in aggregate. Consider three governed financial actions:
Each action is individually below a $5,000 single-action approval threshold. But if the Chain has a $10,000 cumulative threshold, the third action crosses the workflow boundary. That is the core idea behind Arclasp: Govern the workflow, not just the API call.

​
What Arclasp provides

  • Workflow-level Chains — keep ordered governance context across multiple agent actions.
  • Backend-authoritative policy decisions — governed execution depends on the Arclasp backend rather than a model or local fail-open path.
  • Human approval gates — require_approval pauses the governed flow until an authorized reviewer resolves it.
  • Cumulative governance — evaluate running values such as financial exposure and recorded LLM token-cost estimates.
  • Agent risk controls — registered agents can participate in policy, including Critical-tier approval behavior.
  • Fail-closed execution — if Arclasp cannot provide authority for a governed action, the SDK does not silently allow that action.
  • Evidence and verification — completed Chains can produce tamper-evident Chain Records / receipts with authenticated and bounded public verification surfaces.
  • Framework integrations — custom Python, LangChain, LangGraph, CrewAI, and MCP.

​
Who it is for

Arclasp is designed for teams letting AI agents touch consequential systems: money, production infrastructure, customer communications, databases, or workflows that need a durable governance record. If your application only generates suggestions or drafts that a human already reviews before anything happens, you may not need runtime governance yet.

​
How the control loop works

​
What Arclasp does not claim

Arclasp does not:
  • execute your tools for you;
  • see or control actions that never cross an Arclasp governance boundary;
  • provide universal tool-result capture across every adapter;
  • make your system compliant by itself;
  • guarantee legal admissibility or legal non-repudiation;
  • guarantee exactly-once execution of customer-side tools;
  • replace sandboxing, IAM, code review, or observability.

​
Start here

Installation

Install the Python SDK and optional framework extras.

Quickstart

Run a minimal governed workflow with a Chain and receipt.

Chain-level governance

Understand how governed actions accumulate workflow state.

Limitations

Review the current product boundaries and non-claims.