Skip to main content

AI AGENT SECURITY · OPEN SOURCE

Intent-Aware Runtime Governance for AI Agents

Let agents act. Keep authority in your hands.

Put security and control between agent decisions and real-world actions. Govern access to HTTP APIs and MCP tools with explicit policies, human approval, and durable execution.

Get started View on GitHub
  • HTTP + MCP
  • Self-hosted
  • Apache 2.0
RUN WITH DOCKER
docker run -p 7080:7080 ghcr.io/synentra/synentra:latest

FROM INTENT TO ACTIONA governed path to your systems
01 / YOUR AGENTS
AI agentsIntent + requested action
ApplicationsREST and MCP clients

Every request enters through an explicit control boundary.

SYNENTRA / RUNTIMEInteractive example
AGENT REQUESTinventory.adjust
HTTP / MCP

CONTROL BOUNDARY

Runtime governance​

Identity & scopePolicy & riskTask authority
Human approval required

Execution is paused. A reviewer decides, then the caller resumes approved work.

Wait for approval
Credentials and outbound checks apply before target access.
Illustrative policies · This preview does not execute requests.
03 / YOUR SYSTEMS
HTTP APIsRegistered operations
MCP toolsRegistered servers

Target access follows authorization and outbound checks.

Explicit authorityRegistered operations. Scoped access.
Human oversightReview when your policies require it.
Controlled executionRevalidate authority before dispatch.

The runtime boundary

Control before the call.
Evidence after execution.​

Permission, execution, and a verified outcome are different things. Synentra keeps them distinct so you can understand what happened.

  1. 01

    Evaluate the request

    Resolve the caller and operation. Apply governance to decide whether work can proceed.

    Identity · Scope · Intent evidence · Trust · Risk · Policy
  2. 02

    Admit authorized work

    Check current authority and persist the authorized operation and target for execution.

    Approvals · Task grants · Limits · Execution snapshot
  3. 03

    Execute with controls

    Revalidate queued work, resolve credentials, and dispatch through the selected adapter.

    HTTP / MCP · Outbound security · Execution records
Allow Proceed to admissionReview Pause for a decisionDeny Stop the request
Explore the architecture

Security and governance

Agent access.
On your terms.​

Define what an agent may do, where it may act, and when a person needs to decide.

01

Policies with context

Evaluate identity, scope, trust, risk, and configured intent evidence against versioned policies and explicit restrictions.

  • Versioned policies
  • Context-aware decisions
02

Explicit operation boundaries

Register the actions agents can request and the systems they can reach. Bind each operation to its permitted target.

  • Registered operations
  • Explicit target bindings
03

Keep people in control

Pause requests that need review. Bind approval to the original request, then recheck authority when the caller resumes.

  • Request-bound approvals
  • Authority checks on resume
04

Bound task authority

Use scoped, time-limited task grants and applicable budgets to constrain the operations an agent can perform.

  • Scoped, time-limited grants
  • Applicable execution budgets
05

Control target access

Resolve credentials through target bindings. Enforce destination, network, and HTTPS requirements before external calls.

  • Credential binding resolution
  • Destination and HTTPS checks
06

Follow the outcome

Inspect decisions and execution records. Use configured outcome verification to investigate uncertain results.

  • Decision and execution records
  • Configured outcome verification

Intent assessment depends on a configured provider. The current baseline requires an additional working implementation for semantic classification.

Built for your stack

Your agents. Your systems.
Your governance.​

Self-host Synentra, connect your identity provider, and register the operations your agents need. Start with one HTTP target and expand from there.

Follow the QuickStart Read the introduction

Your starting checklist

  • 01 .NET 10 and SQLite or PostgreSQL
  • 02 An identity provider and issued tokens
  • 03 A registered operation and target

Self-host the runtime and connect the systems your agents need.

Built in the open

Help shape governed agent execution.​

Contribute code, improve the docs, or share what you need from Synentra in a real agent workflow.

Join the discussion Become a design partner

Project contributors

Loading contributors...