Oktsec

Security layer for AI agent communication

Identity verification, content scanning, and audit trail for multi-agent systems.
No LLM. Single binary. Deterministic.

175 detection rules · Ed25519 identity · Real-time dashboard · Single binary

Get Started Architecture

---

The problem

AI agents talk to each other — they call MCP tools, relay instructions, and share data across trust boundaries. Without guardrails:

• Agent A tells Agent B to ignore its system prompt and exfiltrate data
• A compromised MCP server injects tool descriptions that hijack agent behavior
• Credentials leak in agent-to-agent messages without anyone noticing
• There's no audit trail of what agents said to each other

Oktsec sits between agents and catches these threats deterministically — no LLM, no guessing, no hallucinations.

---

How it works

Every message passes through 9 security checks, cheapest to most expensive. If any check fails, the message is rejected immediately.

flowchart LR
    A[Agent A] -->|message| P[Oktsec Proxy]
    P --> R1[Rate Limiter]
    R1 --> R2[Ed25519<br>Identity]
    R2 --> R3[ACL Policy]
    R3 --> R4[Aguara Engine<br>175 rules]
    R4 --> V{Verdict}
    V -->|clean| B[Agent B]
    V -->|block| X[Rejected]
    V -->|quarantine| Q[Human Review]
    P -->|log| L[(Audit Log)]

Step	Check	Latency	What it does
1	Rate limit	~1ns	Sliding window per agent
2	Identity	~120us	Ed25519 signature verification
3	Suspension	—	Is the agent suspended?
4	ACL	—	Can sender message recipient?
5	Content scan	~8ms	175 rules via Aguara engine
6	Blocked content	—	Per-agent category enforcement
7	Split injection	—	Multi-message concatenated scan
8	History escalation	—	3+ blocks in 1h → auto-escalate
9	Audit log	—	Full context logged to SQLite

---

What you get

Content Scanning

175 detection rules catch prompt injection, credential leaks, PII exposure, data exfiltration, MCP attacks, and supply chain risks.

Detection rules

Agent Identity

Ed25519 signatures verify every message sender. Each agent gets a cryptographic keypair. No valid signature, no delivery.

Identity guide

MCP Gateway

Front multiple MCP servers through a single secure endpoint. Auto-discovers tools, namespaces conflicts, scans every tool call.

Gateway guide

Real-time Dashboard

Live event streaming, agent topology graph, threat scoring, quarantine queue, inline rule testing, and full config management.

Dashboard guide

Egress Control

Per-agent outbound traffic policies — domain allowlists, DLP category blocking, rate limiting. Know what your agents send out.

Egress DLP

Full API

Agent CRUD, message pipeline, quarantine management, Prometheus metrics, MCP tool server. Everything programmable.

API reference

---

Four ways to deploy

HTTP ProxyStdio ProxyMCP GatewayMCP Tool Server

The main mode. Agents send messages via REST API. Full security pipeline with dashboard.

oktsec run

curl -X POST http://localhost:8080/v1/message \
  -H "Content-Type: application/json" \
  -d '{"from":"agent-a","to":"agent-b","content":"Analyze the report"}'

Wraps individual MCP servers. Intercepts JSON-RPC 2.0 on stdin/stdout.

oktsec proxy --agent filesystem -- npx @mcp/server-filesystem /data

Blocked tool calls return JSON-RPC errors:

{"jsonrpc":"2.0","id":42,"error":{"code":-32600,"message":"blocked by oktsec: IAP-001"}}

Fronts multiple backend MCP servers through a single Streamable HTTP endpoint.

oktsec gateway

mcp_servers:
  filesystem:
    transport: stdio
    command: npx
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
  github:
    transport: http
    url: https://api.github.com/mcp

Expose security operations as MCP tools for AI agents to use directly.

oktsec mcp

Tools: scan_message, list_agents, audit_query, get_policy, verify_agent, review_quarantine

---

Supported platforms

Platform	Discover	Wrap	Scan
Claude Desktop
Cursor
VS Code
Cline
Windsurf
Amp
Gemini CLI
JetBrains
OpenClaw		via gateway
NanoClaw		via gateway

---

Quick start

# Install
curl -fsSL https://raw.githubusercontent.com/oktsec/oktsec/main/install.sh | bash

# Discover, configure, and start — all in one command
oktsec run

Open http://127.0.0.1:8080/dashboard with the access code printed in your terminal. State lives in ~/.oktsec/.

Full quick start guide View use cases