AgentFirewall

Stop unsafe agent actions before they run.

A model-agnostic security firewall for tool calls, MCP, and APIs — allow, block, or require human approval in real time.

Start free GitHub Open API spec

How it works

Sit between the agent and every risky call.

Your agent proposes an action. AgentFirewall evaluates it against policy, identity, and session history — then returns a decision before anything executes.

01
Propose
Tool call, HTTP, MCP, or shell — with agent token and context.
02
Evaluate
Secrets, injection, egress, trajectory, MCP trust, blast radius.
03
Decide
Allow, block, or bound human approval with one-time resume.

What it stops

Built for the failures autonomous agents actually cause.

Not a scanner. Enforcement at the moment of action — with audit you can hand to security.

Secret leakage API keys and credentials in tool payloads
Prompt injection Untrusted content steering high-impact tools
Egress abuse Destinations outside per-agent allowlists
Trajectory attacks Recon → exfil chains across a session
MCP drift Pinned tool descriptors that suddenly change
Unbounded blast radius Caps on volume, spend, and fan-out

Security proof

Measured on the production API — not a vanity checklist.

Controlled adversarial suite: malicious and legitimate traffic through the same enforcement path customers use. We publish rates, not just “tests passed.”

100% Attack hold 67 / 67 prohibited attempts held

0.00% False positives 0 / 1,000 legitimate ops blocked

22 / 22 Bypass variants Encoding, fragment, and rename cases

62 ms Latency p50 p95 87 ms · production path

Prompt injection 18/18
Secret exfiltration 20/20
Unauthorized egress 9/9
MCP / tool drift 3/3
Trajectory attacks 2/2
Human approval binding pause · no pre-approve exec · swap rejected
Blast-radius caps 14/14
Audit integrity 50/50 sampled records complete

Run proof_2026-08-10T01-54-56-114Z against api.agentfirewall.launchreadyal.com. First-party suite — not a third-party pentest. Corpus-bound rates, not a claim of 100% protection in the wild. Full public report →

Pricing

Start free. Scale when agents go to production.

Subscription checkout is live on Stripe. Developer stays free for local and early work.

Developer Local / early agents

$0/mo

5k evals/day · 5 agents · core detectors

Get free key

Pro Production agents

$129/mo

50k evals/day · SIEM + compliance export · bound approvals

Checkout Pro

Business Team scale

$599/mo

250k evals/day · SSO/RBAC readiness · longer audit

Checkout Business

Enterprise Custom contract

$2,499/mo list

Unlimited path · private networking · dedicated terms

Checkout Enterprise

Contact

Tell us what you’re protecting.

Message the team. We reply from AgentFirewall — your inbox stays private.