Sentinel-IR and Sentinel-IR MCP are open for sponsors (and yes, an AI writes most of the code)

A 5-minute read. One maintainer, two Apache-2.0 packages, zero venture capital, a mildly embarrassing amount of honesty.

The short version

sentinel-ir and
sentinel-ir-mcp now
accept sponsorship. Both READMEs have a Sponsor section, both repos have a
GitHub Sponsor button, and sentinel-ir publishes a
funding.json
manifest that funds and tools can read.

People: github.com/sponsors/jakcmencz. The profile was submitted today and GitHub is still reviewing it, so the link turns into a real Sponsor page once they approve it.
Companies, grants, integrations: write to jakubrojicek11@gmail.com. That works today.

What these two things are, in one paragraph each

sentinel-ir reads a source file and gives an AI agent a small fact sheet
instead of the whole file: risk level, network, filesystem, process and
environment calls, routes, imports and exports, and for Rust also unsafe,
FFI, build.rs and proc-macro behaviour. It covers JavaScript, TypeScript,
Python and Rust. It's static analysis: it never runs the code, and a "low"
risk is a hint, not a safety certificate.

sentinel-ir-mcp plugs that fact sheet into Claude Desktop, Cursor,
Windsurf and other MCP clients as a read-only stdio server. It only opens
supported source files inside the folders you allow, refuses traversal,
symlinks and hard links, has file-size, file-count and time limits, and can
keep an audit log of what it read (paths and hashes, never source).

One analyzer, one shared IR, the same facts in npm, the MCP server and the
browser playground. Adding
a language once adds it everywhere.

The honest part

I'm Jakub, the only maintainer, and I'm not a professional developer. I decide
what Sentinel should do, test it on my own machines and approve every release.
Most of the code is written by an AI coding agent working from my specs.
Every change still has to pass the test suites and CI on Linux and Windows
before it ships.

Some grant programmes don't fund AI-generated work, and I'd rather lose a grant
than pretend. If that's a deal-breaker for you, fair enough. If you think
"one person plus an agent shipping audited, tested releases" is exactly the
kind of thing worth backing, read on.

What it saves you (measured, not vibes)

Every time an agent "just takes a quick look" at a file, you pay for every
token of it, comments, licence header and the 400-line date formatter
included. Sentinel hands the agent the fact sheet instead. Measured on
Sentinel's own source with the o200k_base tokenizer:

file	the whole file	Sentinel's fact sheet
`src/ast-facts.js`	8,439 tokens	120 tokens
`src/sentinel-ir.js`	2,084 tokens	133 tokens
`src/tree-parser.js`	891 tokens	121 tokens
`examples/basic.js`	150 tokens	120 tokens

That's up to about 70 times less on a big file. On a tiny file it saves almost
nothing, because the fact sheet has a floor of about 120 tokens, and we'd
rather tell you than let you find out.

A back-of-the-envelope example (an illustration, not a measurement): your
agents open a file the size of ast-facts.js 200 times a day across the team.
That's about 1.7 million tokens a day you don't send. At an example price of
$3 per million input tokens, that's about $5 a day, or roughly $100 a working
month, for one file. Your repo has more than one file. Your agents are not
shy.

The time part is simpler: a model that reads 120 tokens gets to the point
sooner than one that reads 8,439, and its context window stays free for the
work you actually asked for. It's the difference between asking a colleague
"does this file touch the network?" and making them read it aloud first.

And the part that doesn't show up on an invoice: the agent sees
network: outbound and process: spawn before it trusts a file, instead of
discovering them three tool calls later in a postinstall script.

Why monthly, not one-off

A code analyzer is less like a hammer and more like antivirus definitions.
The day it stops being updated, it starts going stale:

Ecosystems keep inventing new doors. A new HTTP client, a new way to spawn a process, a new Rust build crate. Each one Sentinel doesn't know about is a door the agent walks past without noticing. Version 0.4.4 alone taught it axios, got, undici, superagent and fs-extra.
Languages keep coming. JavaScript, TypeScript, Python and Rust today; Go is next. Because everything sits on one shared IR, every new language shows up in the npm package, the MCP server and the playground at once. You pay for it once and get it three times.
Security work never finishes. The MCP server reads files on your machine for an AI. Path checks, size limits and parser robustness need regular adversarial testing, not a one-time audit framed on the wall.
The platforms move. New Node versions, new MCP clients, new MCP spec revisions. Somebody has to keep setup and doctor working on Monday morning.

A one-off donation is a very kind pizza. A monthly sponsorship is the
kitchen. We're grateful for pizza. The kitchen is what keeps the releases
coming.

For companies the maths is short: if Sentinel saves your team even one
engineer-hour a month in review time or token spend, the $5 Supporter plan
paid for itself before lunch. The Development plan is less than many
companies spend on coffee for the people who'd otherwise read those files.

Where the money goes

what	why
AI agent time	it writes the code, the tests and the release checks
CI and test machines	Linux and Windows CI on Node 20, 22 and 24, plus macOS test runs
Go support	the next language for the shared IR
security hardening	more adversarial tests for the parsers and the MCP path checks

The funding.json lists two plans: Supporter, $5/month (helps cover CI)
and Development, $12,000/year (a year of agent time, test machines, Go
and security work on both packages). One-time sponsorship is fine too.

Where we are today

About 1,150 npm downloads a week for each package (Nov 7–13). That number includes our own CI, so it's an upper bound on real users, not a vanity metric.
sentinel-ir-mcp is in the official MCP Registry as io.github.jakcmencz/sentinel-ir-mcp and on Glama.
Current releases: sentinel-ir 0.5.6 and sentinel-ir-mcp 0.5.6.

Can't sponsor? Still helps

Try it on your own repo: npx -y sentinel-ir-mcp@stable-0.5 setup.
Star the repo on GitLab or GitHub.
Open an issue when it gets something wrong. A false "low" is the bug I want most.
Tell the one colleague who keeps pasting whole files into ChatGPT.

Thanks for reading. If Sentinel ever saves your agent from reading
/etc/passwd, you know where the button is.

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.