Lory

Security review while you write, not three weeks after you ship.

Lory is Lorikeet Security's AI security analyst. This extension puts it in your editor: it reads the file you are working on and marks the lines that carry a vulnerability, as you write them.

Findings land as ordinary editor diagnostics. They squiggle in the gutter, count in the Problems panel, and carry a quick fix. There is no separate panel to check and no scan to remember to run.

What it catches

Injection reachable from user input (SQL, OS command, template, deserialization, path traversal, XXE), hardcoded secrets and private keys, missing authorisation on sensitive handlers, broken crypto and predictable randomness, unsafe DOM sinks, SSRF, insecure direct object references, and unsafe framework, container, and cloud configuration.

It deliberately does not report style, formatting, naming, or performance. A security marker on a style nit teaches you to ignore security markers.

Install

Search Lory in the Extensions view, or:

ext install lorikeet-security.lory

No account, no API key, no configuration. The first time it wants to review something it asks permission, once.

How it works

        you type
            │
            ▼
  debounce (1.5s, configurable)
            │
            ▼
  the active file  ──HTTPS──▶  lorikeetsecurity.com/ptaas/ajax/lory-code-review.php
                                            │
                                            ▼
                                   Lory, running the code-review
                                   and findings-format skills
                                            │
            ┌───────────────────────────────┘
            ▼
  findings stream back, one at a time
            │
            ▼
  anchored to lines · filtered · rendered as diagnostics

Every line sent is numbered before it reaches the model, so a finding points at the statement that causes it rather than a line the model counted to. Findings whose line number falls outside the reviewed range are discarded rather than moved to the nearest line: a marker on innocent code is worse than a marker you never saw.

Reviews are cached by content hash. Editing a comment and pressing save does not spend a request, and neither does switching tabs and back.

Using it

Ignore comments

// lory-ignore  the id is an internal enum, not user input
db.query(`SELECT * FROM t WHERE kind='${KIND}'`);

lory-ignore and lory-ignore-next-line both silence the following line, and either also works at the end of the offending line. lory-ignore-file anywhere in the first 20 lines skips the whole file.

Settings

lory.scanOn, the filters, lory.languages, and lory.exclude are all resource-scoped, so a monorepo can review its services and ignore its generated clients.

What leaves your machine

Reviewing code means uploading it. The extension asks before the first review and does nothing until you agree. Lory: Reset Code-Sharing Consent revokes it.

Sent, per review:

• the text of the file or selection being reviewed
• its base filename, never its path
• its language id

Not sent, ever: any other file, your repository, its git history, its remotes, your credentials, or your workspace layout. The extension reads no file you do not have open, and the request carries no identifier tying it to you or your machine.

Nothing is retained after the review returns. If a path must never be uploaded, put it in lory.exclude and it is skipped before any request is built.

Limits

It reviews a fragment, not a system. It usually cannot see the caller, your middleware, or your ORM's escaping, so it states the precondition and marks its confidence rather than guessing. Filter on lory.minimumConfidence if you want only what is provable from the lines on screen.

It is an assistant, not an audit. It finds defects that are visible in source. It does not prove exploitability, and a clean file is not a certified one. For an actual assessment, Lorikeet runs penetration tests that a human reviews and signs.

Requests are rate limited. Thirty a minute anonymously, keyed to your IP — generous for one developer and tight for a team behind one egress address. If you hit it, raise lory.debounceMs, switch lory.scanOn to save, or add a token (below).

Rate limits and access tokens

Run Lory: Set Access Token and paste the lkmcp_… token from your portal's MCP page.

The token buys throughput and nothing else. This endpoint reads no finding, no engagement, and nothing else belonging to your company, so the token grants no data access here — it only identifies you well enough to give you your own bucket.

It is stored in VS Code's SecretStorage, never in settings.json: settings sync between machines, get committed to dotfile repos, and are readable by every other extension in the window. Lory: Clear Access Token removes it and returns you to the anonymous tier.

Development

npm install
npm run watch      # esbuild, rebuilds on change
# then press F5 in VS Code to launch the Extension Development Host

npm run check      # tsc --noEmit && eslint
npm test           # node --test, no editor required
npm run package    # produces lory-<version>.vsix

The logic worth testing is deliberately free of vscode imports — src/glob.ts, src/sse.ts, and src/types.ts — so the suite runs on plain Node.

Server side

The extension talks to /ptaas/ajax/lory-code-review.php in the lorikeet-security repo, which assembles its prompt from the code context of LorySkillLoader: identity + code-review + response-format-findings. That context carries no sales skills, so this surface cannot answer with pricing or a booking.

Legal

Use this on code you are authorised to review. Findings are advisory and reviewing code with it does not constitute a penetration test, a security audit, or a compliance attestation.

License

MIT. See LICENSE.