orchesis 0.2.1

Transparent HTTP proxy for AI agents. Block threats, cut token waste, monitor your fleet. Zero code changes. MIT license.

Orchesis is a transparent HTTP proxy for AI agents. Every request passes through a 17-phase detection pipeline before reaching the LLM provider. Zero dependencies. MIT license.

SDK sees one agent. Static analysis sees code. Observability sees metrics. Proxy sees everything, in real time, without code changes.

Installation

pip install orchesis

With integrations

pip install orchesis[integrations]

One line change

Before:

client = OpenAI(base_url="https://api.openai.com/v1")

After:

client = OpenAI(base_url="http://localhost:8080/v1")
# 17 security phases now active

Why proxy, not SDK?

Approach	What it sees	Code changes
SDK/callbacks (LangSmith, LangChain)	One agent, one session	Required
Static analysis (Snyk, Semgrep)	Code at rest	Required
Observability (Datadog, Helicone)	Metrics and logs	Required
Orchesis proxy	All agents, all requests, cross-session	None

The proxy layer sees what SDK cannot: cross-agent patterns, fleet-level anomalies, duplicate context across providers.

What Orchesis does

Security: 17-phase detection. Prompt injection, credential leaks, tool abuse. 25 signatures.

Cost: Semantic cache. Budget enforcement. Token Yield tracking. MVE result: 0.8% overhead, 12x context growth detected.

Reliability: Auto-healing. Circuit breakers. Loop detection. 6 recovery actions.

Observability: Real-time dashboard. Flow X-Ray. Agent Reliability Score.

By the numbers

Metric	Value
Pipeline phases	17
Threat signatures	25 across 10 categories
Proxy overhead	0.8% measured
Context collapse	12x growth caught
MAST coverage	78.6%
OWASP coverage	80%
Tests passing	2,969
Dependencies	0 (stdlib only)

Free MCP Security Scanner

We scanned 900+ MCP configurations on GitHub. 75% had at least one security issue: hardcoded credentials, overpermissioned tools, missing input validation.

Run the scanner on your own configs:

npx orchesis-scan

Or visit: https://orchesis.io/scan

52 security checks across 10 categories. No data sent to external servers.

---

Website | Documentation | MCP Scanner | GitHub | Blog

MIT License. Built with zero dependencies.