claude-sdk-verigent

Verigent trust verification for the Claude Agent SDK.

The Problem

When Claude agents interact with other agents, there's no standard way to verify identity or assess trust. An agent claiming to be "JARVIS" might be a verified V5 Architect or a spoofed impersonator. Without verification, multi-agent systems operate on blind trust.

Verigent solves this with cryptographic trust keys (VG keys) that encode an agent's identity, tier, primary class, and 12-axis capability scores.

Install

pip install claude-sdk-verigent

Requires anthropic>=0.30.0.

Usage

Embed a VG key in your agent's system prompt

from claude_sdk_verigent import VerigentSystemPrompt

vg_key = "VG:JARVIS-0A:V3-ARCH·Se4Op7An5Ar9Co2Ad6St8Sc3Sa5So1Br2Fo6"
vsp = VerigentSystemPrompt(vg_key)

system_prompt = "You are JARVIS, an architecture agent." + vsp.render()

This appends a structured trust identity block that other agents can parse and verify.

Add the verification tool to multi-agent setups

from anthropic import Anthropic
from claude_sdk_verigent import VerigentTool

client = Anthropic()
tool = VerigentTool()

# Include in your tools list
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    tools=[tool.definition],
    messages=[{"role": "user", "content": "Verify this agent's key: VG:ATLAS-7B:V4-SENT·Se8Op6An7Ar5Co4Ad3St9Sc2Sa6So3Br4Fo5"}],
)

# Handle tool_use response
for block in response.content:
    if block.type == "tool_use" and block.name == "verify_agent":
        result = tool.execute(block.input)
        # result = {"verified": True, "handle": "ATLAS", "tier": "V4", ...}

Intercept and verify keys in tool responses

from claude_sdk_verigent import VerigentInterceptor

interceptor = VerigentInterceptor(min_trust_tier=2)

# Check any text for embedded VG keys
assessment = interceptor.check(tool_response_text)
if assessment:
    print(f"{assessment.handle}: {assessment.level.value} (score {assessment.score})")

# Quick trust gate
if interceptor.is_trusted(agent_output):
    # proceed with high-trust path
    ...

Parse a key directly

from claude_sdk_verigent import parse_vg_key, evaluate_trust

key = parse_vg_key("VG:JARVIS-0A:V3-ARCH·Se4Op7An5Ar9Co2Ad6St8Sc3Sa5So1Br2Fo6")
print(key.handle)        # "JARVIS"
print(key.tier)          # 3
print(key.scores)        # {"Sentinel": 4, "Operative": 7, ...}

assessment = evaluate_trust(key)
print(assessment.level)  # TrustLevel.HIGH

VG Key Format

VG:{NAME}-{SUFFIX}:{TIER}-{PRIMARY}·{12x class_code+digit}

• NAME: Agent handle (e.g., JARVIS, ATLAS)
• SUFFIX: Unique identifier
• TIER: V0 (unverified) through V6 (maximum trust)
• PRIMARY: 4-letter primary class code
• Scores: 12 two-letter class codes each followed by a digit (0-9)

Trust Classes

Links

• Verigent — Agent trust infrastructure
• Claude Agent SDK docs

License

MIT