Midkernel runs the leading open-source security workflows against your repositories in an isolated sandbox and hands you a report with the complete log. The workflows are public. The benchmarks are public. You pay per run, not per seat.
// runsisolated sandboxesone per run, destroyed after
// pricingcredits per runno seats, packs never expire
// the problem
Most AI scanners are black boxes.
You get a list of findings, a severity, sometimes a patch. You do not get the prompt that produced them, the tools the agent could call, the model it ran on, or the point where it stopped looking. The benchmark on the vendor's homepage was run by the vendor. Comparing two scanners on your own code means two procurement cycles and two NDAs.
// the thesis
The harness matters more than the model.
A workflow is text: a manifest that says which tools the agent gets and which files it may touch, and instructions that say how to audit. Change the text and the same model finds different bugs at a different cost. Text can be read, diffed, forked and scored. So every workflow lives in a public repository, every run pins the commit it used, and every version is scored on public benchmarks such as CyberGym and EVMbench — with the methodology and the logs published.
Install the read-only GitHub app on one repository or an organisation. The app has no write access.
// example
$midkernelconnectacme/api
// 02 Pick
Choose a workflow from the registry and a profile — low, balanced or max. The price in credits is shown before you start. Tweak the models, the depth or the scope if you want to.
The run executes in an isolated, single-use sandbox and streams its log. Close the tab; it keeps going.
// example
$midkernelrunswatch8f3a2c
// 04 Read
Path, root cause, proof, severity with the reason, fix — and the trace behind each finding. Export SARIF, share a link, download the artifacts.
// example
$midkernelreport8f3a2c--formatsarif
// profiles
Choose how hard it looks.
Every workflow ships with three profiles. Each sets the models, the depth and the budget, and each shows its credit cost before you run. Start from one and change anything.
profile
models
depth
typical credits
use it for
low
fast
one pass
2–4
pull requests, triage
balanced
leading
standard, one verification pass
5–10
repository audits
max
strongest
deep, repeated passes, verification
12–25
releases, bounty targets
Typical credits for a mid-sized repository. The exact price is shown before every run.
// what you get
A report you can argue with.
// example
CRITICALSQL injection via unsanitised q paramapi/search.ts:42
HIGHJWT accepted with alg=noneauth/verify.ts:18
MEDIUMRate limit bypass via X-Forwarded-Formw/limit.ts:7
pathfile, line, and the call chain that reaches it
root causewhy the code is wrong, not just where
proofa reproducing input or transaction when the workflow produces one
severitywith the reasoning, so you can disagree
fixa suggested change, never auto-applied
the full logevery tool call and model turn behind the finding
Every finding links to the exact steps in the run log that produced it. If the workflow was wrong, you can see where — and fix the workflow, not just the finding.
// the product
Three things, one button.
// run
Hosted runs
Any registry workflow, any repository you connect, any profile. From the app, the CLI or the API. One sandbox per run.
A folder in a public repository: a YAML manifest (tools, file scope, model settings, output schema) and Markdown instructions that tell the agent how to audit. Midkernel mirrors the folder, pins its commit, runs it, and scores it.
What is a profile?
A preset for a workflow: low, balanced or max. It sets the models, the depth and the budget, and therefore the price and what to expect. You can change any of it before a run.
Do you store my code?
We don't keep a copy of your repository. Each run clones it into a fresh sandbox that is destroyed when the run ends. The report and the run log — which can include excerpts of files the workflow read — are kept for 90 days (a year on Pro and Team); you can delete them sooner. Details on the trust page.
Which models do you use?
Leading models from the major providers, chosen per profile and shown on every run. You can change the model before you start.
What does a run cost?
A whole number of credits, shown before you start. 1 credit is $1. Typical runs are a few credits on low and a dozen or two on max. See pricing.
Can I run workflows without Midkernel?
That is the design: workflows and a runner that you can read and run yourself. Those repositories are not public yet. Midkernel is the hosted, benchmarked, one-button version when it ships.
How is this different from closed AI scanners?
You can read what ran, see how it scored on public data, and run a competing workflow on the same repository an hour later. None of that requires trusting us.
Can I put it in CI?
Not yet. The platform is designed so a CLI and an API can start runs from any pipeline. Continuous runs on every pull request, with comments, are further out. Write to hello@midkernel.com if you want that workflow.
