Protect
Stop runaway agents before they become expensive.
Pre-admission hard deny on tool-loop depth, $ caps, RPM, scopes. Freeze the desk when something’s on fire.
My AI agent must never go out of control.
The safety layer between AI agents and the internet.
Tollgate — Protect · Route · Prove. Stop runaway tool loops. Enforce budgets. Survive provider failures. Produce evidence.
- OpenAI drop-in
- Anthropic
- MCP
- n8n
Lane · customer-support
00:00:00Z
// agent loop detected
tool_calls_est20 / 20
Blocked
max_tool_calls · hard deny before next call
✓ budget protected · no runaway spend
// provider failure log
- t+0ms openai · OUTAGE
- t+12ms tollgate · circuit open · re-route
- t+48ms anthropic · fallback admit
- t+91ms RECOVERED · response delivered
// chaos test · simulated outage
10 000requests
9 997recovered
✓ 99.97% resilience · certificate ready
That’s the product. When the agent loops, you get a hard stop. When a provider dies, traffic moves. When you ask “does DR work?” — you have a number.
01 / WHY
Three modes. Not a feature matrix.
You already have agents. You need a gate that says no — and can prove the yes path still works.
Route
Keep the agent running when a provider dies at 3am.
Health-aware routing, circuit breakers, failover — not a hope and a dashboard full of red.
Prove
Don’t assume failover works. Break it on purpose.
Chaos tests, resilience score, certificate scorecard. Evidence for you and for the people who ask.
02 / SIT
Where Tollgate sits
One gate. Every agent lane. Every model behind it.
Not a model catalog. An admission gate.
03 / DESK
Control Room
One glance: safe, expensive, broken — or frozen.
tollgate · local
● protected
94
reliability
$42.18spent
99.8%success
2failovers
⚠ coding-agent approaching budget
| Provider | Health | Latency |
|---|---|---|
| OpenAI | 99.9% | 820ms |
| Anthropic | 99.8% | 1.1s |
| DeepSeek | 97.2% | 2.4s |
Everything the desk needs to stay under control.
Real UI after
tollgate serve →
127.0.0.1:8787/dashboard
04 / WIRE
Point the SDK. Don’t rewrite the agent.
OpenAI-compatible base URL. Protection is on the wire.
- OpenAI SDK base_url
- Anthropic Messages /v1/messages
- MCP stdio
- n8n consumer lane
# no app rewrite — only the endpoint from openai import OpenAI client = OpenAI( base_url="http://127.0.0.1:8787/v1", api_key="support-agent", # or id:secret )
Click code to select · button copies · else ⌘C
05 / START
Cold path. No feature tour.
Protect demo needs no API keys. Full chaos needs keys + ≥2 providers.
- Start Tollgatedocker compose up · or tollgate serve
- Optional keysUser/Key.txt for real chat
- Point the appOPENAI_BASE_URL → :8787/v1
- Set protectionwizard or consumer-budget
- Run provechaos test · certificate
git clone https://github.com/landjunge/tollgate.git
cd tollgate
python3 -m venv .venv && .venv/bin/pip install -e .
./scripts/ten-minute.sh
# or
docker compose up -d --build
open http://127.0.0.1:8787/dashboard
Dashboard: http://127.0.0.1:8787/dashboard
06 / FAQ
Common questions
Short answers for humans and search engines.
What
What is Tollgate (landjunge)?
Open-source AI agent safety layer: budgets, tool-loop hard stops, failover, chaos proof — not a model catalog. Repo: github.com/landjunge/tollgate.
vs gateway
Is it LiteLLM?
No. LiteLLM connects models. Tollgate keeps agents in line and proves DR works.
Name
OpenTollGate or road tolls?
No. Unrelated to OpenTollGate (network payments) and unrelated to highway tolls. This is agent safety software only.
Keys
Need API keys for the demo?
Protect (tool-loop block) works without keys. Real chat and full chaos need User/Key.txt.
07 / GRIT
Self-hosted. Open source. Honest.
Not multi-tenant SaaS. A safety layer next to your agents.
- Self-hosted
- MIT
- Python ≥ 3.11
- OpenAI API
- Anthropic API
- MCP
- n8n
- Docker
- CLI
- HTTP API
- Control Room
Repo URL: https://github.com/landjunge/tollgate