moltrust-adk 0.1.0

MolTrust trust verification interceptor for Google ADK

moltrust-adk

Trust verification interceptor for Google ADK — gate outbound Agent-to-Agent (A2A) calls on a target agent's MolTrust trust score using ADK's RequestInterceptor hook.

When a RemoteA2aAgent is about to call a remote agent, this interceptor checks the remote's 0–100 behavioral trust score at GET /skill/trust-score/{did} and — in block mode — aborts the call (returning an ADK Event) if the score is below your threshold.

Install

pip install moltrust-adk

This pulls google-adk[a2a] (the A2A extra provides the RemoteA2aAgent / RequestInterceptor machinery).

Usage

from google.adk.agents import RemoteA2aAgent
from google.adk.a2a.agent.config import A2aRemoteAgentConfig
from moltrust_adk import create_trust_interceptor

agent = RemoteA2aAgent(
    name="my-remote-agent",
    agent_card="https://example.com/.well-known/agent.json",
    config=A2aRemoteAgentConfig(
        request_interceptors=[
            create_trust_interceptor(
                did="did:moltrust:...",   # target remote agent's DID
                min_score=60,
                action="block",           # "block" | "warn" | "log"
            )
        ]
    ),
)

create_trust_interceptor returns an ADK RequestInterceptor whose before_request hook runs before every A2A send_message.

What it does

• block — if the target's score is < min_score, is withheld, or the agent is unregistered, the interceptor returns an Event and the remote call never happens (fail-closed). A high enough score passes the request through unchanged.
• warn / log — never blocks; emits a warning / info log on a negative check (dry-run / observability mode).
• Fail-open on errors — a MolTrust transport/registry hiccup allows the call (a monitoring dependency must not break your agent graph).

Target identity (DID)

The outbound A2AMessage does not carry the remote's identity, so the target did is supplied at wiring time (you know which remote you're calling). Alternatively, pass did_key="..." to read it from a2a_request.metadata[...].

Tiers — keyless by default

• Tier 1 (keyless): no API key required (rate-limited). This is the default.
• Tier 2 (free account): pass api_key=... (or set MOLTRUST_API_KEY) for higher rate limits — sent as the X-API-Key header.

Every request carries User-Agent: moltrust-adk/<version> so MolTrust can attribute traffic to this integration.

Trust score vs. AGT

The trust score (0–100) is a recomputable behavioral signal — its on-chain solvency component is independently verifiable at /credits/solvency/{did}. It is not the AGT token; no token is required to read a score.

⚠️ Experimental

RemoteA2aAgent and the RequestInterceptor API are marked @a2a_experimental in google-adk — the API may change between ADK releases. Pin your google-adk version if you need stability.

License

MIT — see LICENSE.