Introducing Strathon: the open-source AI agent firewall
An AI agent decides what to do at runtime. It reads a prompt, picks a tool, fills in the arguments, and calls it — querying a production database, sending a customer email, moving money, invoking an internal tool over MCP. That decision is made fresh every time, which means every action is a place where a prompt injection, a hijacked goal, or a plain bug turns into a real-world incident. Nothing sat between the model and the tool to stop it.
Strathon is an open-source firewall for exactly that gap. It sits between your agent and its tools, evaluates every call against policies you write, and blocks the dangerous ones before they execute, inside your agent process, in under a millisecond.
Three lines
You instrument your existing agent. Nothing else about your code changes.
from strathon import Client, instrument
client = Client(api_key="stra_...")
instrument(client, frameworks=["crewai"])
# a matched block policy raises StrathonPolicyBlocked
# before the tool ever executesThe firewall emits OpenTelemetry-native spans for everything it sees, so the dashboard shows you a full trace waterfall of agent behavior, showing which tools fired, what each cost, and which policies matched.
Policies are just CEL
Rules are written in CEL (Common Expression Language), the same expression language Google uses in Firebase, Kubernetes, and Envoy. If you don't know CEL, the dashboard ships 12 OWASP-aligned templates and an AI prompt that turns plain English into a rule.
attrs["gen_ai.tool.name"] in ["delete", "drop_table", "send_email"]Set that policy to require_approval and any matching call pauses until a human signs off. Set it to block and the call never executes.
Self-host it for free
The whole thing is open source. No license keys, no phone-home. Clone the repo, run Docker Compose, and open the dashboard:
git clone https://github.com/strathon/strathon.git
cd strathon
docker compose up -dRead the quickstart to get your first blocked injection in five minutes.