Why Your AI Coding Agent Needs a Doctor’s Filing Cabinet. Sentinel Ir
Imagine walking into a doctor’s office. You hand over your ID, and the nurse walks over to a massive metal cabinet. Inside, thousands of patient records are thrown together in one giant, unsorted pile. No folders. No alp
Imagine walking into a doctor’s office. You hand over your ID, and the nurse walks over to a massive metal cabinet. Inside, thousands of patient records are thrown together in one giant, unsorted pile. No folders. No alphabetical dividers. No insurance codes.
To find your file, she has to pull out every single sheet of paper, read it from top to bottom, and put it back—until she finally stumbles upon your name.
In the real world, no doctor works like that. A real clinic organizes records by insurance provider code, then alphabetically by last name. When Mr. Novak walks in with insurance code 205, the nurse pulls his folder in three seconds flat.
It’s boring. It’s obvious. And it saves hundreds of hours.
Yet, when we give AI coding agents access to our software repositories today, we still hand them an unsorted pile of paper.
Links
Playground: https://sentinel-ir-playground-dc30c8.gitlab.io/
GitHub: sentinel-ir · sentinel-ir-mcp · sentinel-ir-playground
GitLab (main): sentinel-ir · sentinel-ir-mcp · sentinel-ir-playground
Glama: https://glama.ai/mcp/servers/jakcmencz/sentinel-ir-mcp
MCP Registry: io.github.jakcmencz/sentinel-ir-mcp
npm: sentinel-ir · sentinel-ir-mcp
Issues are welcome on GitHub or GitLab. Bug reports with a file that fools
the analyzer are the best kind.
The Hidden Cost of AI-Generated Code
We are entering an era where AI agents generate billions of lines of code every day. Repositories are growing faster than humans—or even the agents themselves—can comfortably read.
When an AI agent (in Claude Desktop, Cursor, or Windsurf) tries to understand a project today, its default behavior is brute force:
- List the directory.
- Open massive source files.
- Read thousands of lines of implementation details, comments, and boilerplate just to find where a single function is defined or how two modules connect. Every time it does this, two things happen: • Context bloat: The model’s attention gets diluted by noise, leading to hallucinations and missed edge cases. • Token & latency waste: You pay for tens of thousands of tokens just for the agent to "look around" the room. As codebases scale, raw file reading stops working. You don't want your agent reading every single sheet of paper in the cabinet just to find Mr. Novak.
Building the "Boring" Infrastructure: Sentinel IR
When I started building sentinel-ir-mcp, I wasn't trying to invent a flashy AI wrapper. I wanted to build the doctor’s filing cabinet for codebases.
Instead of forcing an LLM to parse raw JavaScript, TypeScript, and Python files over and over again, Sentinel IR uses a fast, local Rust engine to extract a compact Structural Intermediate Representation (IR)—a deterministic map of symbols, signatures, exports, and relationships.
Think of it as the index card on the front of the drawer:
• Code 205 → Novak: The agent queries the structural IR first, sees the exact shape of the codebase in milliseconds, and only opens the raw source file when it actually needs to read or edit the implementation.
• Local-first & deterministic: Powered by Rust and running locally over stdio via the Model Context Protocol (MCP), it doesn't guess, it doesn't upload your code to a third-party cloud, and it works offline.
Why "Boring" Reliability Matters More Than Hype
A filing cabinet is only useful if the drawer doesn't jam when you pull it.
In real-world engineering, developers don't work on pristine laptops with open internet. They work behind corporate proxies, custom TLS certificates, weird Windows paths with non-ASCII characters, and restricted environments. If an MCP server crashes the moment it hits a proxy or a self-signed certificate, it becomes another piece of fragile hype.
That’s why most of the work behind sentinel-ir-mcp went into the unglamorous parts: cross-platform determinism, strict environment diagnostics (sentinel-ir-mcp doctor), and adversarial testing across macOS, Windows, and Linux.
Looking Ahead
Ten years from now, when AI agents are maintaining codebases orders of magnitude larger than today's monorepos, nobody will want agents brute-forcing raw text files. The winners in AI tooling won't just be the smartest models—they will be the quiet, reliable indexing layers underneath them.
Sometimes, the best way to move forward is to organize the cabinet.
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.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.