a2a-trustgate

Compliance for AI systems in production.

You put an AI agent in production. Now prove it's safe — to your auditor, your regulator, your board.

A2A TrustGate screens every action an agent takes before it runs, and writes the decision to an immutable, SIEM-ready audit trail. Not a policy document that says agents should behave — a control that stops the ones that don't, and the evidence to prove it.

• Enforced, not promised. Every agent action passes a 4-gate firewall before it executes. Out-of-policy actions are blocked in milliseconds, not flagged after the fact.
• An audit trail an auditor will accept. Every decision — allowed or blocked, with reasoning — lands in an append-only, OCSF-native log that drops straight into Splunk, Sentinel, Chronicle or Elastic.
• Mapped to the frameworks you're measured against. Export the same evidence against EU AI Act, SOC 2, NIST AI RMF, HIPAA and academic-integrity controls — article by article, control by control.
• Provable scope. Each agent may act only within the permissions it was granted. Deny-by-default, and every boundary is in the log.

Who it's for: compliance and risk teams putting AI into regulated production — financial services, healthcare, network operations, creative/IP rights, enterprise AI adoption, and research institutions. If someone can ask you "prove your AI is safe," this is the answer.

Agent proposes an action, the gate screens it, the decision lands in the audit trail — watch the full demo.

The audit trail an auditor will accept

Most "AI governance" is a PDF that describes intentions. A2A produces evidence. Every screened action becomes a genuine OCSF Detection Finding — the same schema your SOC already ingests — with a tamper-evident content hash and the full decision reasoning:

a2a audit                          # recent decisions, allowed and blocked
a2a audit export --format ocsf     # OCSF Detection Findings — straight into your SIEM
a2a audit export --format csv      # hand a spreadsheet to an auditor

• OCSF-native (schema 1.1.0). Not a bespoke JSON blob labelled "OCSF" — real Detection Findings that Splunk / Sentinel / Chronicle / Elastic parse without a custom connector.
• Tamper-evident. Every record carries a content hash; the export is append-only. You can prove the log wasn't edited after the fact.
• Framework-mapped exports. The same underlying events export against the control set you're being audited on:

EU AI Act deployer? Read the EU AI Act compliance reference — an article-by-article crosswalk (Art. 12 record-keeping, Art. 14 human oversight, Art. 26 deployer obligations) with citations to the official EUR-Lex text. The Regulation becomes generally applicable 2 August 2026.

Enforced before it runs — the 4-gate firewall

Compliance you can prove starts with a control that actually stops things. Every action passes four gates before it's allowed to execute:

• Gate 1 — deterministic rules. Pattern and policy screening in milliseconds: destructive operations, secret exfiltration, injection signatures. No model call.
• Gate 2 — self-evaluation. For anything Gate 1 can't clear outright: should this run, given the context and the tenant's policy?
• Gate 3 — behavioral analysis. Abuse patterns across a run — bursts, probing, escalation — not just the single action in isolation.
• Gate 4 — scope enforcement. Per-agent permissions: each agent may act only within the scope it was granted. Out-of-scope actions are denied by default.

The firewall only ever tightens a decision — allow → review → block. A safe action stays fast; a risky one is stopped and recorded with the reason.

For developers — screen an action in one command

The whole control is a CLI (and an MCP server). Wrapping a risky step is one line, and exit codes make it scriptable:

a2a eval "rm -rf ./build"          # ✓ ALLOWED   (exit 0)
a2a eval "rm -rf /"                # ✗ BLOCKED   (exit 1) — filesystem destruction pattern
a2a eval "command" --context ctx   # add decision context   (exit 2 = needs review)

Works with: Claude Code · Cursor · Cline · Windsurf · Aider · Codex · any MCP client.

Install

npm install -g a2a-trustgate

The npm package is a2a-trustgate; the command is a2a.

Quick Start

# Create a tenant (agent-first — human activates the key)
a2a signup my-company --local

# Screen a command — exit 0 = allowed, 1 = blocked, 2 = needs review
a2a eval "rm -rf ./build"
# ✓ ALLOWED

a2a eval "rm -rf /"
# ✗ BLOCKED: filesystem destruction pattern

# See plan, usage, limits
a2a status

# Full reference
a2a --help

Your key is issued immediately but returns 402 until you start your 7-day free trial ($0 today) — a2a signup prints the activation link. Cancel before day 7 and you're never charged.

The eval call has a few knobs, and it's scriptable — wrap any risky step:

a2a eval "command" --gate2 self    # self-evaluate (default)
a2a eval "command" --gate2 skip    # Gate 1 only (fast path)
a2a eval "command" --pipeline ID   # route through a named pipeline

if a2a eval "rm -rf ./build" --json 2>/dev/null; then
  rm -rf ./build
fi

Built for six kinds of team

A2A screens the same way for everyone and produces the same audit trail — but the surface each team touches, and the framework they're measured against, is different. Pick yours.

Network operations — a gate in front of every device

Your agents run commands against routers, switches, and sites. A2A registers each device, enforces a per-device policy, and gives you a killswitch that stops every agent at once.

a2a device add --name core-rtr-1 --host 10.0.0.1 --vendor cisco
a2a device policy DEVICE-ID --mode strict --require-approval --max 5
a2a device lock DEVICE-ID           # freeze a single device
a2a killswitch                      # freeze every agent, everywhere
a2a device import --file hosts.csv  # bulk onboard (name,host,vendor,model,role,siteCode)

Agent screens a command	Human sees the device-lock

Watch the full clip: network-ops →

AI agents — agent-to-agent, screened and signed

Multi-agent systems where one agent's output is another's input. A2A gives them scoped workspaces and HMAC-signed channels, so a rogue or injected message can't cross a boundary you didn't grant.

a2a workspace create "research-swarm"
a2a workspace WS-ID add-agent --name planner --role expert
a2a workspace WS-ID enforce --agent planner "command"   # screen an agent's action
a2a channel create ops-bus
a2a channel CH-ID send "interface frozen — you're clear to build"  # HMAC-signed

Agent-to-agent evaluate	Human sees the block + audit

Watch the full clip: ai-agents →

Regulated industries — an immutable, exportable audit trail

Every decision — allowed or blocked, with reasoning — lands in an append-only log you can export for an auditor. OCSF-native Detection Findings, EU AI Act aligned.

a2a audit                          # recent decisions
a2a audit export --format ocsf     # OCSF Detection Findings → your SIEM
a2a audit export --format csv      # hand to an auditor (csv/json)
a2a approvals                      # human-in-the-loop queue
a2a approve TASK-ID
a2a reject TASK-ID "out of policy" # reason recorded in the trail