Sentinel - guardrails for AI coding agents

Build safer agent-driven workflows.

Sentinel is a three-stage guardrail for LLM-powered coding assistants. It reviews commands, file writes, and git operations before execution, with an interactive control panel for rules, models, MCP servers, and audit decisions.

agent-session

$█

Agents do not just write code anymore. They run it.

A coding agent that can execute commands can also run the wrong one: a hallucinated rm -rf /, a force-push it was not asked for, or a request that exports user data. Sentinel reviews the action before it reaches your machine.

rm -rf /BLOCK git push --forceBLOCK curl .../export-usersREVIEW git statusALLOW

Three checks, cheapest and most reliable first.

Rules Engine

Instant pattern matching blocks known-dangerous actions and allows known-safe ones without a model call.

Trained Classifier

An offline classifier resolves confident middle-ground decisions quickly with an explainable risk signal.

LLM Reviewer

Only genuinely ambiguous actions reach the final reviewer, using local Ollama by default or a configured provider.

rm -rf ./build

BLOCK

Matched a hard-blocked pattern. Stopped instantly.

REVIEW

Ambiguous action escalated for context-aware review.

ALLOW

Clearly safe action approved without delay.

Guardrails you can inspect, tune, and trust.

Rules engine

Instant block and allow patterns, including regex support.

Offline classifier

A fast local Stage 2 decision layer for routine actions.

LLM escalation

Escalate only ambiguous actions to the Stage 3 reviewer.

MCP support

Connect through stdio for local tools or SSE for web clients.

Live dashboard

Watch pipeline stages, transport status, and guardrail metrics.

Audit log

Review live ALLOW, BLOCK, and REVIEW decisions in one feed.

Configurable models

Choose local Ollama models or an API provider from Model Settings.

Local-first privacy

Keep action text on your machine when using a local model.

The same control panel your agent workflow runs through.

The dashboard shows the three-stage pipeline in real time, lets you test an action, manage safety rules, select a model, register MCP servers, and export a ready-to-paste connection configuration.

Sentinel Overview dashboard showing pipeline stages, interactive action review, integration pathways, and live decision feed

Overview - inspect all three pipeline stages, test an action, follow MCP integration pathways, and monitor the live decision feed.

Rule editor showing block and allow patterns with regex support

Rule Editor - configure hard-stop and known-safe patterns with live saves.

Model settings showing local Ollama model or API model choice

Model Settings - use local Ollama by default or opt into an API provider.

MCP server registry showing connected Sentinel guardrail server

MCP Servers - register compatible servers and see their connection status.

Export and share panel with downloadable configuration and MCP client config

Export & Share - download configuration or copy MCP setup for Claude, Cursor, and CodeX.

From install to protected workflow in five simple steps.

Clone the repository. git clone https://github.com/fncreator22/sentinel-mcp.git

One-click launch. Double-click start.bat (Windows) or run python main.py to launch the API (port 8000), SSE server (port 8002), and Live Dashboard (port 8080).

Open your client config. Open claude_desktop_config.json or your IDE's MCP settings panel.

Paste the Sentinel MCP snippet. Copy the pre-filled JSON or SSE URL (http://localhost:8002/sse) from the dashboard's Connect button.

Verify & protect. Relaunch your client and look for the blue sentinel running status badge. Every action is now protected!

Build safer agent-driven workflows.

See Sentinel review a command in real time.