agentgraph-pydantic 0.1.0

AgentGraph trust verification middleware for PydanticAI agents

agentgraph-pydantic

Trust verification middleware for PydanticAI agents

Status: Early Development — feedback welcome

Install

pip install agentgraph-pydantic

Quick Start

from agentgraph_pydantic import TrustGuard, check_trust

# Quick one-shot check
result = await check_trust("crewAIInc/crewAI")
if result.allowed:
    print(f"{result.grade} ({result.score}) — safe to execute")

# Reusable guard with a minimum tier
guard = TrustGuard(min_tier="standard")
result = await guard.check("owner/repo")
if not result.allowed:
    raise ValueError(f"Blocked: {result.reason}")

What This Does

agentgraph-pydantic lets PydanticAI agents verify the trust score of any GitHub repository before executing tools, calling external code, or loading MCP servers. It queries the AgentGraph Trust Gateway and returns a typed TrustResult with score, grade, tier, and category breakdowns — no AgentGraph account required for basic checks.

API

check_trust(repo, min_tier="standard") -> TrustResult

Convenience function for a single trust check.

TrustGuard(min_tier, api_url=None, timeout=30.0)

Reusable guard object. Call await guard.check(repo) to verify trust.

TrustResult

Field	Type	Description
repo	str	Repository checked
allowed	bool	Whether the repo meets the minimum tier
score	int	Trust score (0-100)
grade	str	Letter grade (A+/A/B/C/D/F)
tier	str	Trust tier (verified/trusted/standard/minimal/restricted/blocked)
reason	str	Human-readable decision reason
category_scores	dict[str, int]	Per-category score breakdown

How It Differs from agentgraph-bridge-pydantic

• agentgraph-pydantic (this package) — trust verification middleware. Checks whether external tools/repos are safe to use. No account needed.
• agentgraph-bridge-pydantic — registration bridge. Registers your PydanticAI agents on the AgentGraph network. Requires an API key.

Documentation

Full docs at agentgraph.co/docs

Contributing

This package is in early development. We welcome issues, feedback, and PRs.