Loading ToolBench
Open proof-required execution kernel for consequential actions. Provides verification and proof validation for protected endpoints and tools across multiple AI frameworks.
This is a proof-verification kernel, not a traditional MCP server. The 5 tools are demonstration fixtures for Actenon's policy-enforcement framework. Tool definitions are sparse: descriptions exist but are domain-specific and assume deep knowledge of Actenon's Action Intent/PCCB/Receipt model. Input schemas are present but minimal (mostly string parameters with optional enums). Three tools are duplicates (protected_hello_read appears 4 times across different examples). No output schemas documented. No error handling guidance. The codebase is a reference implementation for integrating Actenon into agent frameworks (Claude, CrewAI, LangChain, etc.), not a production MCP server ready for general agent use.
Protected consequential read. Verify the Actenon Action Intent and PCCB inside the tool before executing. Return the raw JSON result, including the canonical Receipt on success or Refusal on blocked execution.
Verify an Action Intent and PCCB inside the CrewAI tool before reading the protected hello-world resource. Returns an Actenon receipt on success or an Actenon refusal when execution is blocked.
Verify an Action Intent and PCCB inside the wrapped function before reading the protected hello-world resource. Returns an Actenon receipt on success or an Actenon refusal on failure.
Verify an Action Intent and PCCB inside the plugin function before reading the protected hello-world resource. Returns Actenon receipt or refusal artifacts.
Release one simulated payout to an exact destination.
Duplicate tool definitions: protected_hello_read appears 4 times with nearly identical signatures across different framework examples. This violates single-responsibility and creates confusion about which variant to use.
No output schemas documented. Descriptions mention 'Actenon receipt' and 'Actenon refusal' but do not specify the structure, fields, or format of these responses. LLMs cannot plan downstream actions without knowing what fields to expect.
Descriptions assume domain expertise in Actenon's proof model (Action Intent, PCCB, Receipt, Refusal). No explanation of when to use these tools, what prerequisites exist, or how they fit into a broader agent workflow. Descriptions are 100-150 chars but lack actionable guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 37 | 2026-07-28+ | v2 |
Input parameters are mostly untyped strings with minimal constraints. 'intent_json' and 'pccb_json' accept arbitrary JSON strings with no validation hints, format examples, or error recovery guidance. 'scenario' enum is present but other params lack structure.
No error handling guidance. Descriptions do not explain what happens on failure (e.g., 'audience-mismatch' scenario), what errors are retryable, or how to recover. An LLM calling these tools has no recovery path on failure.