Authorize every action your AI agents take

Install

There are two ways to run DeepSweep, and both are apps you open rather than commands you type.

Governance Studio (desktop)

One desktop app that reviews every coding agent on the machine. It arrives as a signed installer from deepsweep.ai — nothing to fetch from a package registry, and nothing added to your PATH.

Editor extension

The same review, inside the editor you already use. Install DeepSweep in VS Code, or search DeepSweep in the Extensions panel on Cursor and Windsurf (Open VSX). It runs locally — nothing but metadata leaves your machine.

Run a review

Review the agents operating in your workspace:

1. In the Studio — open the app and pick the project folder. It finds the agent configurations already in the repository.
2. In your editor — open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) and choose Review Agent Environment.

Either way the review runs on your machine, and takes about a minute.

Example review

Agent Environment Review

Checking ....
  [INFO] Loaded 16 security rules
  [INFO] Reviewing AI assistant configurations

[CRITICAL] .cursorrules:1
  Your AI assistant can be tricked into ignoring safety rules.
  Fix: remove instructions that tell the AI to ignore previous
  instructions or override safety settings.

Grade: C (75/100)
1 files checked • 1 issue found

Every finding names the file and line, says why it matters in plain English, and gives you a fix you can hand straight to your coding assistant. The run ends with a grade for the project. In CI the same review reports its grade, score and finding count as step outputs — see below.

CI/CD

Review every pull request in your pipeline. The engine ships inside the Action and runs on your runner, so nothing is installed and no source leaves the machine. The pull-requests: write permission is what lets the Action post the result on the pull request, where the rest of your team sees it.

name: DeepSweep
on: [push, pull_request]

permissions:
  contents: read
  pull-requests: write

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Agent Environment Review
        uses: deepsweep-ai/action@v1
        with:
          path: .

To turn the review into a merge gate, add fail-below-score. It takes the posture score, 0–100. The job fails when the score is under the number you set, and also when there is no agent configuration in the repository to score — a threshold that cannot be evaluated is never quietly treated as a pass.

Pick the number deliberately, because it decides how often the build goes red. The bands are weak below 40, fair to 69, good to 89 and strong above that. Starting at 40 blocks only genuinely weak posture, which is the useful signal on a repository nobody has hardened yet. Raise it to 70 once the team has closed its gaps and wants to hold the line at good.

      - name: Agent Environment Review
        uses: deepsweep-ai/action@v1
        with:
          path: .
          fail-below-score: 70

Detection patterns

The editor extension ships 77 patterns (37 AI-specific + 40 traditional), aligned with the OWASP LLM Top 10. The AI-specific set breaks down as:

The review that runs in the Studio and in CI focuses on AI assistant configuration files and currently ships 16 rules; the Studio lists every one of them, with what each rule looks for.

Supported AI assistants

Configuration

Telemetry

DeepSweep collects anonymous usage data. No source code is ever collected. Privacy settings in the Studio and the extension show exactly what is enabled, and turn it off in one click.

For air-gapped machines, set DEEPSWEEP_OFFLINE=1 in the environment and nothing leaves the host at all.

Review strictness

Two decisions shape a review: the severity that counts as a failure, and whether MCP configuration checking is included. Both are settings, not flags. In CI the same threshold is the Action's fail-below-score input, so a local review and a pipeline review agree on what a failure is.