Runtime Gateway for AI Agents. Block threats. Cut token waste. Monitor your fleet. One config change.
SDK sees one agent. Static analysis sees code. Proxy sees everything — in real time, without code changes.
pip install orchesis# Before:client = OpenAI(base_url="https://api.openai.com/v1")# After — one line change:client = OpenAI(base_url="http://localhost:8080/v1")# ↑ 17 security phases now active
17-phase adaptive detection. Prompt injection, credential leaks, tool abuse, delegation chain attacks. 33+ signatures across 10 categories.
# Blocked: prompt injection via issue delegation# Phase: Injection Shield (phase 4) + Crystal Alert# Detection: 96% explicit patterns, 0 false positives# Severity: HIGH → request terminated
Context compression saves 80–90% tokens in growing-context sessions. Semantic cache. Thompson Sampling model routing. Per-request budget enforcement — not per-heartbeat.
# Tokens: 4,847 → 612 (87% saved)# Cost: $0.043 → $0.006# Method: context compression + cache hit# Loop detected at call #3 → saved $55-150
Auto-healing with 6 recovery actions. Loop detection at call #3 — saves $55–150 per incident. 450× faster than heartbeat-based checks. MAST & OWASP compliance mapping.
# Loop detected: 3 identical tool calls in 60s# Action: Circuit breaker → 429 Too Many Requests# Saved: $55 (vs $150 at next heartbeat check)# Speed: 450× faster detection
Real-time dashboard, runs locally. Fleet-level correlation: which agent did what, and why it cost so much. Independent audit log — tamper-resistant, outside your orchestrator.
# Fleet: 5 agents monitored# Cross-agent data flow: Agent A → Agent B detected# Cost: Orchesis $4.50 vs reported $0.00# Independent audit: 12,847 requests logged
Live Demo
Security zones show agent health at a glance. Click any agent to inspect.
npx orchesis-radar for your real fleetOrchesis sits as a transparent HTTP proxy between your AI agents and their LLM providers. One config change — set the base URL to localhost:8080 — and every request passes through a 17-phase pipeline. No SDK integration. No code changes. No vendor lock-in.
How it works
base_url = "http://localhost:8080/v1" # one line change
Why proxy, not SDK?
Why agents fail in production
Orchesis closes all four with one config change.
Agent loses the thread. Repeats work. Contradicts itself mid-task.
No trace, no replay. You can't know what went wrong or why.
$847 bill from OpenAI for one overnight run. No limits, no alerts.
Agent called APIs it shouldn't. One incident — banned from production.
One config change. Zero code rewrites. All four problems solved.
By the numbers
Free · No signup · Browser-based
100+ checks. CVE database. OWASP mapping. Paste your config, get a report.
5 questions. Instant grade A+ to F. Know where you stand in 30 seconds.
How much are agent loops and context bloat costing you?
All 10 risks explained. Real examples. Detection methods. Fixes.
Mathematical foundations
What can and cannot be detected in AI agent security — proven mathematically.
We don't just monitor your agents. We prove exactly where monitoring ends — and where your other defenses must begin.
Works with your stack
If it speaks OpenAI-compatible API — Orchesis works with it.
Integrate into your workflow
npx orchesis-scannpm · orchesis-scan →
uses: poushwell/orchesis@main with: fail-on: 'high'GitHub · mcp-scan action →
repo: poushwell/orchesis hooks: - id: orchesis-mcp-scanpre-commit · hook →
How Orchesis compares
| Criteria | Generic Gateway | LLM Router | Agent Platform | Orchesis |
|---|---|---|---|---|
| Understands MCP/A2A | ✗ | ✗ | ✗ | ✓ |
| 17-phase security | ✗ | ✗ | ✗ | ✓ |
| Fleet correlation | ✗ | ✗ | partial | ✓ |
| Formal security proofs | ✗ | ✗ | ✗ | ✓ |
| Zero dependencies | ✗ | ✗ | ✗ | ✓ |
| Open source (MIT) | varies | some | ✗ | ✓ |
| Transparent proxy | ✗ | partial | ✗ | ✓ |
| Zero code changes | ✗ | ✗ | ✗ | ✓ |
| Self-hosted | ✗ | ✗ | ✗ | ✓ |
| No telemetry | ✗ | ✗ | ✗ | ✓ |
Why this matters now
of agentic AI projects will be canceled by 2027 due to inadequate risk controls
stars on Paperclip in 2.5 weeks. AI agent fleets are here. No dedicated security layer exists for them yet.
Runtime Gateway for AI Agents market in 2026, growing to $2.8B by 2030.
EU AI Act enforcement begins. Audit trails and incident reporting required.
Orchesis is ready. Are your agents?
From the blog
Bare shell execution is the most common MCP tool interface. Here's why that's dangerous.
Read more →INCIDENTLiteLLM supply chain attack: Trivy → KICS → PyPI. Full chain analysis.
Read more →INCIDENT$47,000 from an agent loop. 43,175 restarts. 2.5 years of data wiped.
Read more →Open source · MIT License
Now you can see every one.
Works whether AI wins or loses.
Open Source · MIT License · Zero Dependencies