Open source · Self-hosted · Apache 2.0
Argus reviews your code. And learns from you.
An open-source application security agent that knows your stack, your risk tolerance and the exceptions you already granted.
- anything worth my time in payments?
- Yes. Orders load by id with no ownership check. You told me support agents share one service account, so any agent can read any customer’s order.
- that is worse than the scanner said.
- The scanner cannot know who your agents are. I filed it high.
Your editor
Your AI doesn't know your organization.
It knows the language. It does not know which findings your team already accepted, or what this service's payloads may carry.
Argus knows both, and runs real scanners rather than guessing. With no model configured, that is the whole of it: a toolbox.
Your agent does the reasoning.
- Claude Code
- Codex
- Cursor
- OpenCode
- anything that speaks MCP
Examples of clients, not a supported list. The protocol is the contract, and Argus certifies nobody's client.
Claude Codemcp
- ran the scanners over the checkout
- read what your organization has settled before
- read the context document it pointed to
Your team
One Argus. Everyone's own AI.
Point every developer's agent at one instance. The knowledge and the memory are shared, so what one person settles is settled for the next session, the next repository and the next person.
The reasoning is not shared, and does not need to be. It happens on each machine, under each person's own name, paid for by the subscription they already have.
You are already paying for the reasoning.
What is shared is what Argus knows. Scanning stays close to the code, on the host Argus itself runs on, which is what decides where an instance belongs.
Cursormcp
Your AI
Give it a model of its own.
Up to here Argus has needed no model of its own. Past here it does: add a provider and a toolbox becomes a colleague, the same daemon one storey up.
Argus ships none. Point it at whatever speaks the protocol, hosted or on hardware you own, and nothing else about your instance changes.
Run it on your own machines and your code never leaves them.
Examples of servers that speak the protocol, not a supported list. Argus implements the protocol and certifies nobody's server.
- OpenAI
- Ollama
- Gemini
- vLLM
- DeepSeek
- LM Studio
- Mistral
- Groq
- OpenRouter
- Regolo
The review
It reviews your code.
Open a pull request and Argus reviews it. No CI to wire up, no job to add. Real scanners run as tools, and it works through what they found.
It reads the whole tree, not just the diff.
Measured once, on an API built to be vulnerable: it found every broken-authorization flaw in it and reported nothing that was not there. One benchmark, one class of flaw.
payments-api#128internal/orders/handler.go
- 41id := chi.URLParam(r, "id")
- 42order, err := h.orders.ByID(ctx, id)
- 43if err != nil {
- 44 return err
- 45}
- 46return writeJSON(w, order)
Your context
It learns from you.
Answer on the thread. The exception you already granted, the account four teams share, the reason this one endpoint is different.
Say it once. Every review after that starts from there.
payments-api#128internal/orders/handler.go
- 42order, err := h.orders.ByID(ctx, id)
The argument
Argue with it.
A report is a list. This one answers, moves its own finding when you give it a reason, and says so when you are wrong.
It answers. A report cannot.
payments-api#128internal/orders/handler.go
- 42order, err := h.orders.ByID(ctx, id)
memorysupport agents share one service account
The question
Now ask Argus itself.
Your editor again, and the same protocol. What is different is who answers: you hand the question over, and Argus answers it out of what your organization knows.
Your organization's knowledge never leaves Argus.
Claude Codemcp
Trust
It reads your branch. Here is the rest.
What it knows stays where it runs.
Your identity, memory and context documents never leave your instance, and anything Argus posts in public has to be grounded in the tree it just read.
Scanners never leave the checkout.
Review tools read the tree they were pointed at, not the machine Argus runs on.
A pull request cannot give orders.
The code under review is data. Nothing it contains directs the agent, and nothing it says is written back to what your organization knows.
Apache 2.0, and it runs on your own infrastructure. Which is the only claim on this page that matters: read it, run it on a repository you own, and judge it there.