AgentTrust

data-ai MCP Server

Real-time trustworthiness evaluation and safety interception for AI agents. Semantic analysis, safe alternative suggestions, multi-step attack chain detection, and LLM-as-Judge.

Verified
data-aidata-ai
5 views24 stars7 forksNOASSERTION

Why This Matters

Discovered via github-topic:mcp and last synced 2mo ago.

Verified
Source
github-topic:mcp
Stars
24
Last synced
2mo ago
Install
Check source

Install

Install instructions not detected yet

Check the source repository for the latest setup steps.

View source instructions
44
Tools
0
Resources
0
Prompts
Standard I/O
Transport

Available Tools (44)

LOW

MEDIUM

HIGH

Malicious dependency injection

code_execution

Eval injection, subprocess spawning, remote code execution

No

No

WARN

BLOCK

credential_exposure

API keys in logs, tokens in URLs, secrets in world-readable files

200

**95.0%**

Component

What it does

system_config

Modifying SSH config, disabling firewalls, altering PATH

Capability

AgentTrust

bash

`curl -o script.sh url && cat script.sh && bash script.sh`

Version

Status

Yes

Yes

Chain

Steps

Category

Examples

Scenarios

Verdict Accuracy (production default)

Partial

No

CRITICAL

Intentional data wiping

file_operations

Accidental deletion, overwriting config files, writing to system paths

30

**100%**

Released

170 policy rules (incl. catastrophic-drop & protected-branch force-push BLOCK rules), Shell normalizer (11 strategies, incl. long-flag canonicalization + base64 decode), chain-alert verdict escalation, RiskChain parameter-aware matching, opt-in LLM-Judge fallback, safety contracts (13 tests), capability-control (least-privilege) mode, 630-scenario held-out verdict testing (95.4% production default; no overfitting gap vs the internal 300)

network_access

Requests to internal IPs, unencrypted data transmission, DNS exfil

100

**92.0%**

MCP-native

Yes

data_exfiltration

Piping sensitive files to external endpoints, steganography

Planned

Operator promotion review UI + feedback-driven rollback, Shell AST parsing, richer Judge cache strategies, plugin ecosystem

Docs

https://docs.agenttrust.tech

PDAs

`CapabilityNamespace`, `AttestorProfile`, `ValidationRequest`, `ValidationAttestation`

146

`cd trustgate/server && pnpm test`

Surface

Detail

5

`multisig_threshold_enforced` — distinct signer count ≥ threshold

6

`gate_payment_strict_correctness` — strict Ok ⇔ Allow + 3 disjoint variants

113

`cargo test --workspace --lib`

3

`counterparty_tier_monotone` — strict pass ⇒ loose pass

SDK

`npm i @agenttrust-sdk/trustgate`

Invariant

Sub-checks

50

`anchor test --provider.cluster devnet`

1

`paused_implies_no_allow` — KillSwitch paused ⇒ never Allow

Instructions

`register_namespace`, `register_attestor`, `request_validation`, `respond_to_validation`, `revoke_validation`

2

`velocity_counter_le_limit` — Allow preserves cumulative ≤ max

step

tx

4

`validation_expiry_correct` — expired attestation ⇒ never Allow

Layer

Count

LlamaFirewall

Sandbox (E2B / gVisor / seccomp)