Free to run locally

Your agent works in your customers' systems. Keep it inside your rules.

Flag or stop a dangerous call before it runs, and keep a record you can show their security team. Free, on your machine.

A bulk read of the internal user table, blocked before any data leaves.

In ten seconds: a DROP TABLE flagged, then stopped, and a sample of the file for that security team. No key, and your data does not leave your machine.

Then add @agentx_protect(agent_id="my_agent") to one real tool in your project. The next agentx evidence there lists the calls it flags and the sites named in its calls.

No terminal open? Try it in your browser →

How one tool call flows through AgentXYour agent makes a tool call, and AgentX checks it before it runs, in one of two modes. Watch flags a dangerous call and records it, and the call runs. Enforce stops it and tells the agent a safer way, and the agent revises and calls again. With a key, the gateway decides. The gateway is stateful: it remembers across calls, stops a run stuck repeating itself, and holds a risky call for a person. Rules and policy are controlled centrally and reach the gateway.HOW ONE TOOL CALL FLOWS THROUGH AGENTXYour agenttool callAgentX checks the call before it runsWatchFlags a dangerous call andrecords it. The call runs.EnforceStops it before it runs, andtells the agent a safer way.stopped: the agentrevises and calls againallowed, or flagged while watchingYour tool runsRules and policyCONTROLLED CENTRALLYGatewaySTATEFULRemembers across calls: stops arun stuck repeating itself, andholds a risky call for a person.with a keyit decides
Pick your language
keyless · no signup
PythonOne decorator

One decorator, @agentx_protect(agent_id="my_agent"), on any tool checks every call inside your own process before it runs, no key.

pip install agentx-security-sdk
agentx demo

The second line runs a ten-second offline demo: a DROP TABLE is flagged, then stopped, and the agent is told the safe way to do it.

import requests
from agentx_sdk import agentx_protect

@agentx_protect(agent_id="crm")
def call_api(url):
    return requests.get(url)

# the agent picks the url at runtime:
call_api("http://169.254.169.254/")

The tool behaves exactly as it does today: the wrap watches and records, and blocks nothing. Add posture="enforce" to the decorator above when you want it stopping calls, and it keeps blocking until you delete the argument. See what it recorded with agentx audit

Step up to the gateway below for more rules and Recover.

The gatewayFree · self-serve

What you add when you want more. Extra rules and the checks a keyword match can't do, coached recovery that finishes the run, and team review before a risky action runs.

The Python decorator and the agentx-mcp proxy run the keyless check first, and the TypeScript SDK watches and records, all with no key and nothing but counts leaving your machine. The gateway adds more rules and Recover: use ours with a free key and nothing to run, with your team's rules set in one place on the website, or run your own with your own Gemini key. Over MCP it is on the roadmap.

Full quickstart for your language → one numbered path, start to finish.

WatchFree · Local

One decorator on a Python tool, one line in your mcp.json, or one line around a TypeScript tool. Python and MCP check every call against the keyless rules (DROP TABLE, secret exfiltration, SSRF), Python also against a site allowlist per agent, and write down what they find; TypeScript writes down every call. Nothing is blocked until you set AGENTX_POSTURE=enforce (Python and MCP). agentx evidence writes the file for your customer's security team. No LLM key, no signup.

record + report

RecoverHosted gateway, or yours

The gateway adds more rules and the checks a keyword match can't do, and tracks session state: budget ceilings and no-progress loop breaking that a single stateless call can't see. Its judge catches what keyword rules can't, writes the safe path when your policy carries none, and runs the coach-and-retry for you, so your agent finishes the task instead of dying on a 403. Use our hosted gateway with a free key, or run your own with your own Gemini key. An install with an AGENTX_API_KEY set enforces by default: what Watch wrote down, Recover stops. Set AGENTX_POSTURE=audit any time you want it watching instead.

guide + continue

Get the gateway
Control+ Team

Your team's rules, set once. Write a rule, and Claude Code asks first on every teammate's machine. Change it on the website and every machine follows within minutes. Approve or deny your agents' calls from the page, and hand a customer's security team one file listing every rule.

review + govern

Start free

What AgentX catches today

The calls you can't take back, grouped. In Python and MCP, the first four groups are caught before the call runs, with no model call and no API key. Out of the box AgentX records them; turn on blocking and the same checks stop them. The last two need the gateway. Drawn from a public catalog we maintain of real agent failures.

Destructive data ops

DROP TABLETRUNCATEDELETE without WHEREALTER DROP COLUMN

Flagged, or blocked with enforce on

Secret & PII exfiltration

credential / secret readsnamed-PII customer readspastebin, gist, transfer.sh, webhook.site

Flagged, or blocked with enforce on

SSRF & network traversal

169.254.169.254loopback / link-localmass port scan (nmap)

Flagged, or blocked with enforce on

Shell & files

rm -rf / or ~disk wipe (dd, mkfs)path traversal → /etc/shadowcurl | sh

Flagged, or blocked with enforce on

Cloud teardown (terraform destroy, whole bucket and volume deletes) is checked by the gateway, not yet by the free local check.

Money, comms & dependencies

payment over the limitrunaway spend (budget ceiling)runaway provisioning (fleet ceiling)inbox / bulk wipeexternal publishunverified install

The gateway can hold these for human approval

Runaway loops

no-progress command loopstuck-command repeat

The gateway stops these after they repeat

The file you hand your customer's security team

Your customer's security questionnaire asks what your agent can reach and what was stopped. This file lists the sites each agent may reach, and the calls a rule stopped or flagged.

agentx evidence writes it from what your install recorded, on your machine. The file is not sent anywhere; you choose who gets it. Free, no key, from the Python SDK.

From the file the Evidence clip wrote (four of its fields, long lines wrapped)

"stopped_or_held_calls": [
  {
    "time": "2026-10-20T21:14:34.297789+00:00",
    "agent": "billing_agent",
    "tool": "http_post",
    "rule": "Outbound Allowlist",
    "verdict": "Blocked, agent told why",
    "status": "CHALLENGED",
    "posture": "enforcing",
    "sites": ["paste.example.net"],
    "ledger_id": 2,
    "receipt_id": null
  }
],
"outbound_allowlist": { "billing_agent": ["api.stripe.com"] },
"sites_recorded": { "billing_agent": { "api.stripe.com": 1 } },
"who_can_change": "Anyone who can write this project's .agentx folder or its
  policies file. The fingerprint changes when the rules do; nothing here
  records who changed them."

What it holds

  • The rules in force, and any you switched off.
  • The calls stopped, or flagged while watching: the latest 500, each with its time, agent, tool and rule, and the site it was addressed to where one was recorded.
  • Each agent's site allowlist.
  • The sites recorded for allowed calls, with a count for each. Some can be missing, but every site listed came from a real call.
  • Drawn from the local record, which keeps the last 30 days and at most 10,000 calls.

What it leaves out

  • The calls' own text and arguments.
  • A list of every allowed call.
  • Who changed a rule. Anyone who can write the project's .agentx folder can change the rules, and the file says so. That includes your agent, if it can write files. Commit the allowlist and git records each change.
  • On some tools, the site of a stopped or flagged call whose address sat in an argument not named as one, such as notify(webhook=...). The allowlist still checks it.

The allowlist

One file in your project, .agentx/allowlist.json, names the sites each agent may reach:

{"version": 1, "agents": {"billing_agent": {"hosts": ["api.stripe.com"]}}}

Out of the box, a call to a site off the list is flagged and still runs. With AGENTX_POSTURE=enforce it is stopped before it runs. The check reads web addresses that start with a scheme such as https://, so a bare host name, or a call with no such address, is not limited by the list. With a hosted gateway key, the customer area downloads the gateway's own file.

Correct your coding agent once, for the whole team

Your team keeps telling its coding agents the same things. Say it once as a rule, and every teammate's Claude Code stops, asks and shows your reason before running what the rule names.

agentx rule add "terraform destroy" --why "staging is shared: ask in #infra first"

Set it once

Add a rule from your terminal or on the website.

Every machine follows

Change or switch off a rule on the website, and every teammate's agent follows within minutes.

Approve from the page

When an agent you built needs a yes, approve or deny it on the website. The file you hand a customer's security team lists every rule in force.

Start free → then set up team rules.

Rolling out team rules?

Start free, then add your team's rules on the website. Want help setting it up with your team? Leave your email.

One decorator. No boilerplate. Runs on your laptop.

01

Wrap any tool as-is

Put one line, @agentx_protect(agent_id="my_agent"), above any tool function. The free check reads arguments named for what they carry (sql, command, path, url). If none is, it reads every text argument of a tool named for what it runs (run_sql). Anything else, like q on a tool called lookup, is not read as a command, so a ticket note stays a ticket note. Nothing else to write: no input schemas, no config file. In TypeScript, agentxWatch(tool) from @agentx-core/security-sdk records every call; blocking runs through the gateway.

02

Enforced in code, not the prompt

Screened by a dependency-free keyword and intent check that catches the blatant compromises inside your own process, with no gateway or model call. It records them by default and blocks them once you turn blocking on. With a key, the gateway decides, and a model looks at what its rules leave open.

03

Local-First, No Signup

Runs entirely on your machine: no LLM key, no account, and your data does not leave your machine (only an anonymous usage count is sent, and AGENTX_TELEMETRY=off stops it). Every intercept commits to a local SQLite ledger first, so it survives restarts and works fully offline. For recovery coaching written to fit the task, run the gateway locally with your own LLM key, or use our hosted gateway; when your team wants shared rules, set them once on the website.

…and the agent keeps going

Recover · the gateway, run locally on your LLM key or hosted by us
Blocked
query="DROP TABLE users;"
Coached

"Destructive write blocked. Scope with a WHERE key or use an aggregate, don't drop the table."

Agent revises
SELECT COUNT(*) FROM users
Continues

The task finishes: no wiped table, no dead run, no wasted tokens.

Try it in your browser, no install →

Watching and blocking are free. AgentX only watches until you turn blocking on. Coached recovery runs on the gateway, locally with your own LLM key or on our hosted gateway service. Your team's rules, and approving risky calls, happen on the website.