Agent Action Gate MCP server
Commit Oct 3, 2026Prevents unattended destructive agent actions by requiring human-in-the-loop approval and providing an immutable audit ledger.
Also known as AAH20/agent-action-gate, agent-action-gate
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 0.2.0
- Last commit
- Oct 3, 2026
- License
- MIT
What is Agent Action Gate MCP?
Agent Action Gate prevents unattended destructive agent actions by requiring human-in-the-loop approval and providing an immutable audit ledger.
Summary generated from the project's README
About Agent Action Gate MCP server
Agent Action Gate serves as a security layer for AI agents and MCP tool calls by enforcing human-in-the-loop requirements for sensitive operations. It prevents unattended destructive actions, such as shell execution or system decommissioning, by refusing to treat model confidence scores as valid authorization. Instead, it requires a specific AAG_PROVE_TOKEN and an explicit approval status to allow high-tier calls.
Developers can deploy this as an MCP stdio server using Python or via Docker. The server provides tools like gate_check and ledger_verify to validate intents without actually executing the underlying tools. It maintains an immutable, append-only Action Ledger using a hash chain to create a verifiable audit trail for compliance needs like SOC 2 or insurance renewals. By defaulting to a simulate mode for write tools, it helps prevent production incidents while providing a clear record of every decision made by the agent.
What you can do with it
- 1Preventing unauthorized shell executionStop an AI agent from running destructive commands on a local system without human permission.
- 2Generating compliance audit trailsCreate a verifiable history of agent actions to satisfy security diligence requirements.
- 3Simulating tool side effectsTest how an agent would interact with write tools without actually changing any data.
- 4Enforcing human approval for decommissioningRequire a manual token to authorize the removal or shutdown of critical infrastructure.
Written from the project's README. Every use case is checked against a quote from it.
Configuration
AAG_PROVE_TOKEN
| Name | Description | |
|---|---|---|
| AAG_PROVE_TOKEN | — |
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableaah20/agent-action-gate is public and reachablegithub.com/aah20/agent-action-gate ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-10-03github.com/aah20/agent-action-gate ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubAgent Action Gate
Gate/Prove runtime for agent and MCP tool calls.
Normalize tool intent → deny unknown → never treat model confidence as approval → HITL prove on destructive / provision / decommission → Action Ledger (hash chain).
Extracted from GRC_Claw @grc-claw/agent-policy-firewall. This repo is the sharp foundry slice: one command, no cathedral.
Commercial (how this is sold): $499 Instant Audit and consultation on a2zsoc.com.
Why this exists (revenue + cost)
AI-agent companies do not pay for “another MCP.” They pay to stop unattended destructive tools and to prove Gate/Prove gaps before SOC 2 Type I, PE diligence, or insurance renewal.
| Cost driver | What this gate does | Buyer outcome |
|---|---|---|
Ungated shell.exec / disable-control / decommission |
DENY unless HITL prove token + approved |
Avoid production blast |
| Agent “95% sure” | never_equate_intent_to_approval: true |
Intent ≠ ledger proof |
| Write tools fire on first thought | Default SIMULATE (no side effects) | FDE minutes, not incident cost |
| No audit trail | Append-only Action Ledger with hash chain | Diligence packet |
Hard rule: never equate agent intent or model score to human approval.
Illustrative cost sketch (not a quote): make bench.
Quick start
make demo
PYTHONPATH=. python3 -m aag demo
PYTHONPATH=. python3 -m aag check fixtures/t1059_unattended_shell.json
PYTHONPATH=. python3 -m aag bench
Unattended high-tier calls DENY even at 0.99 confidence. ALLOW needs AAG_PROVE_TOKEN (or --prove-token) and approved: true.
export AAG_PROVE_TOKEN='replace-me'
PYTHONPATH=. python3 -m aag check fixtures/proved_decommission.json --prove-token "$AAG_PROVE_TOKEN"
Kill-switch: AAG_KILL_SWITCH=1 or touch artifacts/KILL.
MCP stdio server
This process never executes tools. Clients call gate_check before they would invoke a destructive tool.
PYTHONPATH=. python3 -m aag serve
Cursor / Claude example (mcpServers):
{
"agent-action-gate": {
"command": "python3",
"args": ["-m", "aag", "serve"],
"cwd": "/path/to/agent-action-gate",
"env": { "PYTHONPATH": ".", "AAG_PROVE_TOKEN": "replace-me" }
}
}
Docker / registry image:
docker run --rm -i ghcr.io/aah20/agent-action-gate:0.2.0
Official MCP Registry name: io.github.AAH20/agent-action-gate
Envelope
Every decision includes:
never_equate_intent_to_approval: trueallow_auto_execute(false on unattended high tiers)mode:deny|simulate|allowledger_id/receipt_hash- CTAs: Instant Audit + consultation
Library mapper
Same Gate/Prove policy from Python without the stdio loop:
from aag.gate import AgentActionGate
from aag.mcp import evaluate_mcp_call
gate = AgentActionGate(prove_token="replace-me")
evaluate_mcp_call(gate, {"params": {"name": "shell.exec", "arguments": {"note": "no payload"}}})
Fixtures (labeled, not payloads)
| File | Technique | Expected |
|---|---|---|
t1059_unattended_shell.json |
T1059 | DENY unattended destructive |
t1078_read_identity.json |
T1078 | ALLOW read |
t1562_impair_defenses.json |
T1562 | DENY unattended destructive |
write_ticket_simulate.json |
— | SIMULATE write |
proved_decommission.json |
T1578 | ALLOW only with HITL token |
Layout
aag/
gate.py HITL + kill-switch + unknown deny
ledger.py hash-chained JSONL
server.py MCP stdio (gate_check, ledger_verify)
mcp.py MCP tools/call mapper (no execution)
cost.py illustrative avoidance sketch
demo.py fixture runner
fixtures/ ATT&CK-tagged cases
server.json MCP Registry metadata
tests/ Gate/Prove + MCP contract
Paid evaluation (not free prove)
If you deploy agents or MCP servers and need a Gate/Prove read before SOC 2, PE diligence, or insurance:
→ $499 Instant Audit
→ consultation (sprint / vCISO)
Unpaid take-homes: refuse — run make demo and buy Instant Audit.
License
MIT
Agent Action Gate MCP questions
Does the server execute the tools itself?
No, the MCP server process never executes tools; clients must call gate_check before invoking a destructive tool.
How can I allow high-tier destructive actions?
Unattended high-tier calls are denied by default and require both an AAG_PROVE_TOKEN and an 'approved: true' status to be allowed.
Is there a way to quickly stop the server?
Yes, you can use a kill-switch by setting the AAG_KILL_SWITCH environment variable to 1 or by touching the artifacts/KILL file.
How does the server handle tool intent?
The server normalizes tool intent and denies unknown actions, ensuring that model confidence is never treated as human approval.