Which AI agent is doing what — and should it be allowed to?

DeepSweep Studio™ answers that for every AI coding agent on your machine, from one desktop app. It works local-first and metadata-first: your source code never leaves your machine.

Add DeepSweep to your editor

Free extension $99 human security audit

1 · Review

Point Studio at a workspace and get an Agent Environment Review in about a minute: which AI agents are configured, what they can reach, and where the owner and boundary gaps are.

2 · Authorize

Turn the review into policy: allow, require approval, or deny — per agent and per action. Deny wins, and anything unspecified requires approval by default.

3 · Audit

Every decision lands in a local audit ledger — who (which agent), what, why, and under which policy — so you can answer for agent activity after the fact.

DeepSweep Studio Review — every agent, every capability, every gap
Review — every agent, every capability, every gap
DeepSweep Studio Authorize — per-agent policy, deny wins
Authorize — per-agent policy, deny wins
DeepSweep Studio Audit — the local tamper-evident ledger
Audit — the local tamper-evident ledger

What using it looks like

  1. Install Studio (macOS or Windows — one app for all agents, below).
  2. Open it on a project folder. The first review runs read-only and offline.
  3. Read the review: agents found, tool access, boundary gaps ranked by severity.
  4. Apply the suggested policy, then tighten it: approve-by-default is the starting posture; grant specific agents specific permissions as you trust them.
  5. Keep shipping — the audit ledger records every agent decision locally, and CI can gate on the same review so gaps never merge silently.
MCP Security Gateway

Not yet available to download. It ships with Studio.

Runtime governance for AI agent tools

The gateway will come with the Studio desktop app. It is designed observe-first: it records and explains MCP tool calls routed through it, and enforcement turns on only when you say so. What you can use today is the free review in your editor. The example below shows the kind of MCP findings it reports.

Signal Detect emerging concerns
Assess Analyze risk severity
Policy Generate guardrails
Enforce Apply your policy per call
Observe Monitor effectiveness
Refine Improve detection
Gateway
Example: what the free review reports
[FAIL] Unrestricted MCP Tool Access
[WARN] MCP Auth Disabled
[WARN] Remote MCP Server
[INFO] Sensitive Env Vars in MCP Config

Tool Call Review

Built to check MCP tool calls routed through it against your policy before the call is forwarded.

Approval-Gated Actions

A call your policy marks for a human is refused with a single-use approval token bound to that exact agent, action, resource, and argument set. A human approves it out of band, then the agent retries.

Rug-Pull Circuit Breaker

Tool descriptions are pinned on first sight. A changed description trips a breaker that refuses the call until a human re-approves the tool.

Evidence Ledger

Allowed, denied and held tool calls are written to a hash-chained local ledger with inclusion proofs. Records carry hashes of payloads, never your code.

DeepSweep Studio™

One app to govern every AI coding agent

Review what your agents can reach, authorize what they're allowed to do, and keep an audit trail — Cursor, Claude Code, Copilot, Windsurf and the rest, governed from a single desktop app. No per-editor setup.

DeepSweep Studio — per-agent authorization policy, where deny wins
Authorize — per-agent policy, deny wins. See every screen
LIVE policy v3

Every action, decided before it runs

    …and written to evidence you can check

    TREE HEAD …

    Select an entry to recompute the head from its proof.

      —decisions
      —gated or denied
      —p95 decision

      Timed in your browser and amortised over 5,000 evaluations per row. Browsers deliberately blunt their clocks, so anything under about 0.2 µs is below what this page can measure.

      Real decisions from the real policy engine, and a real evidence ledger — running here, in your browser. Select any entry to verify its proof.

      Early access — first builds go out to this list. All builds and checksums See how it works