atlas-pca-crewai

CrewAI guard for Proof-Carrying Authority (PCA): make every CrewAI tool call proof-carrying and policy-gated. A tool only runs when the agent carries a valid, in-plan, resource-correct proof for exactly that action. The adapter delegates all verification to atlas-pca and reimplements no cryptography.

Status: 0.4.1, unaudited.

What is tested

The default test suite runs without the framework, against the shared conformance proofs. A second suite, make test-real (needs the framework installed), runs against the real framework and is pinned to CrewAI 1.15.26. It covers a real BaseTool and @tool, and a real Crew kickoff driven by a scripted offline LLM: a tool call runs only under a valid proof, and is refused (never executed) without one, with a tampered proof, or for the wrong audience. Each denial test checks that the tool body never ran, and checks the refusal reason wherever the framework surfaces it.

Not covered: other framework versions (only the pinned one is exercised), live model providers (the models are scripted and offline), and any property of the underlying atlas-pca verifier beyond what it already documents. This package has not been independently audited.

Install

pip install atlas-pca-crewai

This installs atlas-pca>=0.4.1,<0.5 as a dependency.

Use

Entry points: pca_tool / guard_crew / PcaToolGuard.

from atlas_pca_crewai import pca_tool

See the module docstrings in atlas_pca_crewai for the full API, and the PCA specification for the model. The verifier applies the full check order, including grant_ref_bound (the signed grant_ref must equal the root capability id), as of atlas-pca 0.2.0.

Links

• Source: https://github.com/Atlas-Authorization/pca-python
• Specification and docs: https://github.com/Atlas-Authorization/pca

License: MIT