
Sponsor Sentinel IR development here: https://github.com/sponsors/jakcmencz?o=syg&sc=r
A three-minute case for giving AI agents a library card, not the keys to the building.
Latest project updates: https://dev.to/jackymencz/sentinel-ir-mcp-updates-0510-0513-5dhp
Verdict
Source (GitLab, Apache-2.0): https://gitlab.com/jakubrojicek11/sentinel-ir
npm: https://www.npmjs.com/package/sentinel-ir
Playground (runs in your browser): https://sentinel-ir-playground-dc30c8.gitlab.io/
MCP Server on npm: npmjs.com
https://glama.ai/mcp/servers/jakcmencz/sentinel-ir-mcp
9:00 AM. The library is open.
A customer asks the librarian for a book.
“Where is The Practical Guide to Production?”
“I don't know,” says the librarian. “It was here a minute ago. The AI has been reorganizing.”
“Again?”
“It says it is improving the layout.”
Books move. Labels disappear. The catalogue becomes fiction. And when the customer asks where something is, the librarian can only shrug.
Would you trust that library?
That is what broad, unnecessary permissions can turn an AI coding agent into: a guest hired to read the catalogue, but handed the keys to rearrange the whole building.
A library card, not a master key
Most coding tasks do not require an agent to rewrite, delete, or freely explore everything it can reach. First, it needs to understand the project: what files exist, what they do, and where attention is needed.
That is the idea behind Sentinel-IR and Sentinel-IR MCP.
Sentinel creates a compact structural view of code and gives the agent a scoped, read-only way to inspect permitted project folders. The MCP server does not write project files, execute code, or open network connections. It helps the agent find the right shelf before asking it to read every book.
Malicious notes can still hide inside a repository. OWASP identifies excessive tools, permissions, and autonomy as risks. Sentinel reduces exposure to raw untrusted text, but does not claim to eliminate prompt injection. The rule is simple: repository content is data, not the agent's new boss.
The receipt matters
The original Sentinel idea was bounded visibility and read-only access. Audit records and SHA-256 evidence came later, strengthening that principle.
With audit logging enabled, Sentinel records the tool, path, result, and file hash, without recording source contents. Current releases also return SHA-256 fingerprints for analyzed files, helping teams check whether a file changed after inspection.
A hash is useful evidence, not a magical force field against log tampering.
Why does this matter to companies? A 2026 OneTrust survey of 1,200 senior decision-makers found that 87% encouraged AI-agent use, but only 47% said clear governance, oversight, and controls were in place.
When security asks, “What could the agent access? What did it inspect? Can we verify it?”, “Trust the agent” is not much of an answer. Tools that cannot explain their boundaries or provide useful evidence may struggle to pass enterprise review.
Give the AI a library card. Let it read what it needs. Keep the keys with the people responsible for the building.
The librarian should never have to say:
“Sorry. The book was here before the AI got promoted.”
Explore Sentinel: "Source" (https://gitlab.com/jakubrojicek11/sentinel-ir) · "MCP server" (https://www.npmjs.com/package/sentinel-ir-mcp) · "Playground" (https://sentinel-ir-playground-dc30c8.gitlab.io/)
Research and standards
- "OWASP: Excessive Agency" (https://genai.owasp.org/llmrisk/llm062025-excessive-agency/) — risks of excessive tools, permissions, and autonomy.
- "OneTrust / Sapio Research: 2026 AI-Ready Governance Survey" (https://www.onetrust.com/resources/onetrust-2026-ai-ready-governance-report/) — 1,200 senior decision-makers; agent adoption is outpacing governance.
- "Check Point Research: AI Security Report 2026" (https://research.checkpoint.com/2026/ai-security-report-2026/) — observed longer indirect prompt-injection payload detections rose roughly fivefold from March to May 2026.
Independent Engineering Projects
Built as independent engineering experiments across AI, software analysis, security, and autonomous systems.
Evaluated by reproducible evidence, not by the number of layers.
Explore the projects:
Sentinel Ecosystem: AutoDoc-Sentinel, Sentinel-IR, Sentinel-IR-MCP and related tools.
Agentix Lite: Independent security and system-observation project.
Find the projects: GitLab repositories | NPM packages
Support the work: ⭐ Star the repositories on GitLab.
Connect: Follow me here on DEV Community or explore the Playground.
Each project is distributed under its own license. Check the corresponding repository or NPM package for licensing terms.
Copyright © 2026 jackymenCZ. All rights reserved. Written for the community.




Top comments (0)