Skip to content
| Marketplace
Sign in
Visual Studio Code>Linters>Agent WatchdogNew to Visual Studio Code? Get it now.
Agent Watchdog

Agent Watchdog

Zorzen Studios

| (0) | Free
Checks the code your AI agent writes: invented packages and calls, broken team rules and risky MCP servers. Every flag shows its source and reason.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

Agent Watchdog

Checks the code your AI agent writes: invented packages, wrong function calls, broken team rules and risky MCP servers. Every flag shows where it came from and why. Most checks need no AI at all.

Works with code from any agent (Copilot, Claude Code, Cursor, Codex, or you): Watchdog checks changed code, whoever wrote it.

Agent Watchdog flags an invented import: the hover on 'express-rate-guardian' says the package doesn't exist (fact check, no AI), suggests express-rate-limit with a one-click fix, and the Changes panel shows the file's trust score.

What it checks

Fact checks (no AI, free)

On the lines that changed in TypeScript, JavaScript and Python files:

  • Imports: is the package installed (node_modules, your virtualenv, the standard library)? Names that look like a typosquat of a popular package are flagged. With registry lookups on, a package that isn't installed is also checked against npm or PyPI.
  • Calls and named imports: does the function exist in the installed type definitions (.d.ts, Python sources and .pyi stubs), and does the call pass a valid number of arguments? The hover shows the correct signature.
  • Python names: from idna import encodee, json.dumpz(...), helpers.slugifyy(...) and from .missing_module import x are flagged when the module (installed, standard library or your own) doesn't define the name, with a "did you mean". Modules that create names at run time (__getattr__, globals(), lazy loaders, compiled builds) are skipped rather than guessed.

Registry checks are opt-in (zorzen.watchdog.registryLookup, Watchdog asks once) and send package names only. Typosquat detection and "did you mean" suggestions use bundled name lists from npm-high-impact by Titus Wormer (MIT) and hugovk/top-pypi-packages. When package.json, a lockfile or a virtualenv changes (for example after npm install), changed files are checked again.

Squiggles follow the severity: ▲ critical (error), ◆ serious (warning), ○ minor (information).

Team rules from .watchdog-rules

A YAML file in the repo root (and optionally in package folders, which inherit from the root). Rules are hot-reloaded on save; parse errors go to the Agent Watchdog output channel and the Rules tab.

  • Exact rules (no AI): import boundaries, banned packages, banned patterns, required file headers. They compile to deterministic checks.
  • Plain-English rules (Pro), such as "Every API route checks the user is logged in", are judged by an AI model: your Copilot models through VS Code first, your own Anthropic or OpenAI key as fallback, and skipped with a notice if neither is available.
version: 1
rules:
  - id: UI-BOUNDARY
    type: exact
    kind: import-boundary
    from: ["ui/**"]
    disallow: ["api/**"]
    severity: serious
    description: ui/ may not import from api/
  - id: AUTH-01
    type: plain
    rule: Every API route checks the user is logged in
    applies_to: ["api/**"]
    severity: serious

The Rules tab lists the repo's .watchdog-rules (exact rules NO-LOG, DB-ONLY and NO-MOMENT, and the plain-English rule AUTH-01); in the editor, a route that imports stripe is flagged by DB-ONLY: "Routes don't call Stripe directly; go through src/payments".

The rule wizard (Pro) turns a sentence into a rule: it picks an exact rule when it can, a plain-English rule when it must, shows the YAML, tests it on the open file and adds it.

The rule wizard: "Use zod instead of joi" becomes an exact banned-package rule (no AI), shown as YAML with a button to test it on the open file.

When a repo has .watchdog-rules, Watchdog offers once to add itself to .vscode/extensions.json, so VS Code suggests it to teammates who clone the repo.

Suppress with a reason

// watchdog-ignore: AUTH-01 public order-tracking link
router.get('/orders/:id', async (req, res) => {

A reason is required. Fact checks use FACT-PACKAGE, FACT-EXPORT, FACT-ARGS and FACT-TYPOSQUAT. Every suppression is recorded in the audit log. The quick fix Suppress … with reason writes the comment for you.

MCP Auditor (free)

Finds the MCP servers configured for VS Code Copilot, Copilot CLI, Cursor, Claude Code, Claude Desktop, Devin Desktop (formerly Windsurf), Gemini CLI and Codex, in the workspace and in your user settings. For each server it shows the config file, the command or URL, version pinning, which clients load it, and what it can reach (files, network, shell, secrets), labelled inferred.

Flags: tokens written in plain text, unpinned npx/uvx packages, unknown publishers, shell-capable servers, remote servers without auth, broad file access. Fixes are edits to the config file you can undo (for example "Move to ${env:GITHUB_TOKEN}" or "Pin to 1.4.2"). Block removes a server from every client config that loads it, after you confirm.

The MCP servers tab lists three servers from .vscode/mcp.json with their flags; the github server's page shows its command, which clients load it, what it can reach (network and secrets) and the flag "Token written in plain text" with a "Move to ${env:GITHUB_PERSONAL_ACCESS_TOKEN}" fix.

Tool snapshots (rug-pull check). "List tools" starts the server (or connects to a remote one) to read its tool names and descriptions, and compares them with the snapshot you accepted earlier. Starting a server runs its code, so this only happens when you ask for it, only in a trusted workspace, and only after a dialog that shows the exact command line (or URL) and which environment variables or headers it gets. Watchdog sends initialize and tools/list, stops the server after at most 10 seconds and never logs its output. Changed descriptions show as a diff with Block server and Accept change.

Sandbox suggestions (Pro): a minimal VS Code sandbox block for local servers on macOS and Linux.

Trust score

Each changed file gets a score; the status bar shows Watchdog 72% across all changed files. Click it for the breakdown.

The Changes tab: a trust score of 17% for one changed file (1 critical, 1 minor), the "How is this scored?" breakdown by imports, calls and team rules, and the list of flags with their lines.

How the trust score is calculated

Trust score = passed checks ÷ all checks, weighted by severity.

Every check Watchdog runs on changed code (an import exists, a call matches its types, a team rule holds) counts once, with the weight of the severity it would flag:

Severity Weight
▲ Critical 8
◆ Serious 4
○ Minor 2

score = 100 × (weight of passed checks) ÷ (weight of all checks), rounded to a whole number.

  • A file with no checks scores 100.
  • Suppressed flags (watchdog-ignore: RULE_ID reason) count as passed. Every suppression is recorded in the audit log with its reason.
  • While any check fails, the score shows at most 99.
  • The overall score adds up the weights of every changed file's checks. It is not an average of the file scores.
  • The score is plain arithmetic over the checks above. No AI decides it.

The trust score summarises the checks that ran. It is not a guarantee that the code is correct or safe.

Privacy

Agent Watchdog has no backend that receives code.

What happens
Code and AI (SEC-01) Fact checks and exact rules run on your machine. Plain-English rules go straight to your Copilot models (vscode.lm) or to the API endpoint of your own key. Nothing goes to Agent Watchdog.
What is sent (SEC-02) Only the changed snippet and the rules that apply to it, at most 4 KB per request. Never whole files. Before the first Copilot request Watchdog explains this, then VS Code asks for its own consent. The status bar shows each request in flight with the provider and the bytes sent, and the hover says how many bytes were sent for a flag.
Keys (SEC-03) API keys and the licence key are kept in VS Code SecretStorage only. Never in settings, workspace files or logs.
Audit log (SEC-04) File paths, line numbers, rule IDs, scores and counts, kept in VS Code's workspace state. Never code.
Untrusted workspaces (SEC-05) Fact checks only: no rule evaluation, no AI requests, no MCP server is started.
MCP inventory (SEC-06) Read locally and kept in memory. Only the tool snapshots you accept (names, descriptions, schema hashes) are stored, in VS Code's global state on this machine.
Network (SEC-07) No requests during start-up. Opt-in registry lookups send package names only. The licence check sends the licence key only, at most weekly, to the server you configured. AI requests go only to the provider you chose.
Telemetry (SEC-08) None.

Settings that decide where data goes (registryLookup, ai.*, license.verifyUrl) are application-scoped: only your own user setting counts, a workspace can't change them.

Account and thank-you moments

Agent Watchdog says thank you, once each: a Thank you tab when your licence activates, a note when your 14-day trial starts, small milestones at the top of the side bar on every plan (10, 100 and 1,000 flags fixed, your first clean week) and One year with Watchdog. Each shows once, also across your computers with the same licence, and never while a check or a save is running. Paid plans get a Watchdog-colour member tag, a thin ring around the logo and Watchdog Pro in the status bar.

Account (click the member row at the top of the side bar, or run Agent Watchdog: Show Account): your plan and since when, what Watchdog caught for you (counted on this computer only, never sent), Write to the maker (an email to support@zorzenstudios.com), Manage plan, Invoices and your licence key.

Plans

Free Pro Team
Fact checks (imports, calls) ✓ ✓ ✓
Exact rules 5 Unlimited Unlimited
Plain-English AI rules, rule wizard ✓ ✓
Audit log export (JSON, Markdown) ✓ ✓
MCP inventory and flags ✓ ✓ ✓
MCP sandbox suggestions ✓ ✓
Suppression with reason ✓ ✓ ✓
Price (excl. VAT) €0 €9/month or €90/year €15/seat/month, min. 3 (coming soon)

Pro has a 14-day trial that starts locally, with no card and no network call (Agent Watchdog: Start 14-Day Pro Trial). AI runs on your own Copilot or key, so the plans sell features, not AI usage. Bundle with MonoVue Team: €29/seat/month.

Commands

All under Agent Watchdog in the Command Palette.

Command What it does
Check This File Checks the whole active file now (including plain-English rules)
Open Panel Changes, Rules and MCP servers
Show Trust Score Breakdown Opens the breakdown in the panel
Open .watchdog-rules / Reload Rules / New Rule (Wizard)… Team rules
Rescan MCP Configs / List MCP Server Tools… MCP Auditor
Choose AI for Plain-English Rules… Copilot models, your own Anthropic or OpenAI key, or off
Export Audit Log as JSON… / as Markdown… Pro
Recommend Agent Watchdog to Teammates (extensions.json) Adds the workspace recommendation
Start 14-Day Pro Trial / Enter Licence Key… / Remove Licence Key / Verify Licence Now / Compare Plans Plan
Show Log The output channel (metadata only)

Settings

Setting Default
zorzen.watchdog.checkOnSave true Check changed lines on save
zorzen.watchdog.checkPendingEdits true Check unsaved edits as they land (fact checks and exact rules; AI runs on save)
zorzen.watchdog.codeLens.enabled true "▲ Fact check" / "◆ AUTH-01" above flagged lines
zorzen.watchdog.registryLookup false Check packages that aren't installed against npm/PyPI (names only). Watchdog asks once.
zorzen.watchdog.ai.provider auto auto, copilot, anthropic, openai or off
zorzen.watchdog.ai.copilotFamily empty Copilot model family; empty = first available
zorzen.watchdog.ai.anthropic.endpoint / .model https://api.anthropic.com/v1/messages / claude-sonnet-5-5 Used with your own key
zorzen.watchdog.ai.openai.endpoint / .model https://api.openai.com/v1/chat/completions / gpt-6.1-terra Used with your own key
zorzen.watchdog.license.verifyUrl https://licence.zorzenstudios.dev/v1/verify Licence server: receives the product name and key, at most weekly. Empty: a stored key can't be verified, so the plan stays Free or trial.

When checks run

  • On save: the lines that changed since Watchdog first saw the file in this session.
  • On edits that land in an open file: Copilot agent mode and other in-editor agents write unsaved edits into the document, and a file a CLI agent rewrites on disk is reloaded by VS Code; both are checked (fact checks and exact rules) a moment later. VS Code's stable API doesn't say which edits came from an agent, so Watchdog treats all changed code the same. Files that aren't open in VS Code are not watched: open one and run Check This File.
  • Check This File: the whole file, from then on.

Requirements

VS Code 1.95 or later. Plain-English rules need Copilot models or an Anthropic/OpenAI API key.

Support

Questions, bug reports and licence help: support@zorzenstudios.com.

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft