mcp-review

An MCP server that gives a coding agent eyes on the UI it just changed: submit a preview URL, get findings with route, viewport, element ref and a suggested fix, apply them yourself, then ask the server to recheck. It judges and verifies; it never edits code, drives a browser, or opens a pull request — the server is the eyes, the agent is the hands. Everything runs offline with no credentials, and with nothing configured the judgments come from a golden fixture that says so in every payload; one clone and one environment variable swap in a real critique of a real page.

Quickstart

Node 24+ and pnpm 9.15.0 (corepack enable installs pnpm from the packageManager field). No API key, no database, no Docker; pnpm install needs the npm registry once, and after that nothing here opens a connection.

git clone https://github.com/apatureai/bastion.git
cd bastion
corepack enable
pnpm install --frozen-lockfile
pnpm build
pnpm demo

pnpm demo is a real MCP client. It spawns the server as a child process over stdio, completes a handshake, submits a review, reads it back four ways, acts on a finding, rechecks it, and is denied on an unauthorized host — the nine steps shown in the hero above. It ends:

[9] design_review  https://evil.example.org/  (SSRF boundary)
    rejected: DOMAIN_UNVERIFIED (next_action: verify_domain)

Done. 3 findings, 3 recheck outcomes.
Open out/panel.html in a browser to see the review panel.

Success looks like nine numbered steps, 3 findings, 3 recheck outcomes, and two new files: out/review.json (the agent-facing Critique) and out/panel.html (the self-contained review panel).

open out/panel.html      # macOS; use xdg-open on Linux

The server still identifies itself on the wire as apature-mcp-review. That name is part of the MCP handshake and any client configuration that already references it, so the repository rename to bastion deliberately left it alone.

If pnpm demo reports Cannot find module, pnpm build has not been run. If step 2 comes back DNS_TARGET_PROHIBITED instead of a job, something changed LOCAL_RESOLVED_ADDRESS in packages/mcp-server/src/local-server.ts to a non-public address, and that rejection is the SSRF guard working.

Related MCP server: vibe-annotations

What you get

pnpm demo reviews https://preview.example.com/pricing and writes two files. out/review.json is the Critique a coding agent reads: a grade, a findings[] array where each entry carries a dimension, severity, route, viewport, element_ref and a suggestion fix string, plus two envelopes an agent must check before trusting any of that — provenance and coverage. out/panel.html is the same review as a self-contained HTML panel: no scripts, no external requests, evidence embedded as data: URIs. The job_* and rev_* ids are freshly generated, so yours differ from the hero's.

With nothing configured the findings are a golden fixture describing a fictional pricing page, not a judgment of the URL you passed — and the payload says so rather than relying on the startup banner, because the consumer is usually an agent that never sees the banner. The full result contract covers grounding, the model-free measurements block, and how every payload an agent reads — the recheck, the panel action, each finding, each content block — carries the same marker.

Provenance: did anything judge this page?

Every Critique carries a provenance object, on every backend including the offline fixture path, so an agent that never sees the terminal can answer "which engine ran" from the payload alone:

"grade": "unjudged",
"confidence": null,
"provenance": { "model_backed": false, "source": "fixture", "engine": "bastion-fixture", "model": null }

The rule an agent codes against: trust the result only when provenance.model_backed === true and coverage.state is not "nothing". When model_backed is false — the fixture engine, or verdict run with --model mock/--model canned — the grade is the literal "unjudged", confidence is null, the narrative says no model judged the page instead of describing one, and every finding carries "unjudged": true. When it is true, the engine's grade, narrative and findings pass through untouched and provenance.model names the judge. A remote verdict deployment reports null: a real judgment may be behind it, but this process cannot see how that deployment is configured and does not claim to.

Coverage: what did it actually look at?

provenance answers whether a model judged the page; it does not answer what the model judged, and those come apart. Verdict's triage can conclude a deep review is needed and then name no route to run it on: a model really was called (model_backed truthfully true), yet no page was judged, and a result with no surviving findings grades ship by construction. So every Critique also carries coverage:

"grade": "nothing_reviewed",
"coverage": { "state": "nothing", "routes_requested": ["/", "/pricing"], "routes_reviewed": [], "routes_skipped": ["/", "/pricing"] }

state is full (nothing suppressed), partial (a real verdict about a smaller surface; routes_skipped names what it missed), nothing (the grade becomes "nothing_reviewed" and every finding is marked unjudged), or unstated (the engine reported no coverage — never read as "everything was reviewed"). nothing_reviewed wins over unjudged when both apply, and both facts stay in the payload. This is the same rule and vocabulary apatureai/gate uses for its Check Run, so the two surfaces cannot tell different stories about one run.

Usage

MCP annotations are set from the truth: only design_review_get and design_review_panel_action carry readOnlyHint: true, because submit and recheck create metered jobs and cancel terminates one.

design_review_get views

The eyes-not-hands boundary, in code

design_review_panel_action is where the product boundary is easiest to violate and easiest to test. When a reviewer clicks "apply fix" on a finding, the server reads the completed Critique, projects it into fix items (reviewFixItemsFromCritique) where a finding is grounded only if it is localizable (element_ref) and carries a concrete repair constraint (suggestion), and runs a pure reducer (handlePanelAction):

• a grounded finding comes back { "type": "fix", "fix": "..." }, and that fix is for the host to hand to the coding agent;

• an advisory finding comes back { "type": "human_only" }, never an auto-fix;

• a finding from a review nothing judged comes back { "type": "unjudged" } with no fix string at all, because on that path the instruction is fixture text and handing fixture text to a coding agent is precisely the failure this boundary exists to prevent.

"Resolved" is a recheck verdict the service earns, not a status the panel can set.

Connect your own MCP client

The local server speaks MCP over stdio — what Claude Code, Cursor, Codex, and VS Code use for a local server. pnpm demo spawns exactly this command and completes a real handshake against it:

{
  "mcpServers": {
    "apature-review-local": {
      "command": "node",
      "args": ["/absolute/path/to/bastion/packages/mcp-server/dist/local-stdio.js"]
    }
  }
}

The only host the local server authorizes is preview.example.com; every other host is rejected as DOMAIN_UNVERIFIED. Add your own hosts and a critique backend through the environment, with no code change and no rebuild:

{
  "mcpServers": {
    "apature-review": {
      "command": "node",
      "args": ["/absolute/path/to/bastion/packages/mcp-server/dist/local-stdio.js"],
      "env": {
        "BASTION_ALLOWED_HOSTS": "preview.mycompany.com",
        "VERDICT_CLI": "/absolute/path/to/verdict",
        "MODEL_BASE_URL": "https://your-openai-compatible-endpoint/v1",
        "MODEL_API_KEY": "..."
      }
    }
  }
}

Getting real judgments

With nothing configured the findings are a fixture about a fictional pricing page. This replaces them with a real critique of a real page — no credentials from this project, no database, no hosted service. The backend is apatureai/verdict, public and MIT: it launches headless Chromium, captures each route at each viewport, measures the DOM, critiques the render against your repository's own design system, deletes every finding it cannot point at, and writes the EngineReviewResult that packages/mcp-types/src/engine.ts declares and this server consumes. Connecting them is configuration, not code.

1. Build verdict

git clone https://github.com/apatureai/verdict.git
cd verdict && corepack enable && pnpm install --frozen-lockfile && pnpm build
pnpm browser:install     # Chromium for playwright-core, roughly 275 MB downloaded

2. Point bastion at it

export VERDICT_CLI=/absolute/path/to/verdict
export MODEL_BASE_URL=https://your-openai-compatible-endpoint/v1
export MODEL_API_KEY=<your-key>

VERDICT_CLI selects the backend; MODEL_BASE_URL and MODEL_API_KEY are verdict's own variables, passed straight through, so any OpenAI-compatible chat-completions endpoint that accepts images works. Neither half is optional: a key with no base URL, or an explicit VERDICT_MODEL=live with no key, is a startup error rather than a silent fallback to the mock — verdict would otherwise answer a live request with its mock client and report model_backed: true anyway.

3. Review a URL

pnpm review https://preview.mycompany.com/pricing --routes /pricing --viewports desktop

pnpm review is the same MCP client pnpm demo uses, pointed at a target you name on the command line (which authorizes that host for the run, and says so before it starts). With a key set, verdict runs --model live and a model looks at your screenshots; without one it runs --model canned, which captures and measures your page for real but judges nothing, and announces that loudly. The per-run artifact directory under out/verdict/ keeps the screenshots, the DOM geometry map, the measured facts, and the resolved system prompt, so a finding can be checked against what the model was actually shown.

The SSRF boundary makes one narrow exception: a loopback host you name explicitly (localhost, 127.0.0.0/8, ::1) is your own dev server, so it may be plain http. The exception is keyed on the literal host, never on where a name resolves, so a public name pointed at loopback is still refused before verdict is even started — DNS-rebind protection:

$ pnpm review https://localtest.me/

rejected: DNS_TARGET_PROHIBITED
  host localtest.me resolves to a prohibited (loopback) address

A rejected target is the boundary working, not a crash: pnpm review prints the code and reason and exits 1 (it exits 2 only on a usage or engine-configuration error, and 0 on a completed review), so an agent loop can tell a refused target apart from a bug.

What each mode gives you

Every critique-backend variable and the honest list of what is not wired yet — including why design_recheck does not work against either verdict backend — are in docs/configuration.md.

Configuration

None of the variables below are needed by the local server, the quickstart, or the test suite — they configure the Streamable HTTP edge. Dockerfile builds the workspace and runs packages/mcp-server/dist/boot.js, the production composition root: Streamable HTTP transport, bearer JWT verification against an issuer's JWKS, a Postgres application plane, and a signed client for the judgment engine. It fails closed with a readable message when configuration is missing, and it has no mock fallback — a server that answers with fixture judgments must be the local one, explicitly. Nobody operates a public instance, so directory/server.json declares no remotes.

The critique-backend variables (VERDICT_CLI, MODEL_BASE_URL, and the rest) are listed separately in docs/configuration.md; ENGINE_BASE_URL and ENGINE_HMAC_SECRET are shared by both surfaces.

Design notes

• docs/result-contract.md — how grounding, the model-free measurements block, and every agent-read payload carry the provenance marker.

• docs/design-notes.md — who this is for, why a design reviewer breaks the thin-MCP-wrapper assumptions, and the two independent version numbers.

• docs/how-it-works.md — the production request path, what is real versus synthetic offline, and the repository map.

• docs/configuration.md — the full critique-backend variable reference and what is not wired yet.

• docs/roadmap.md — what works today and the known gaps, each with its seam.

Status

The local MCP server, all five tools and five views, target authorization and egress classification, job lifecycle and recheck semantics, the Streamable HTTP edge (OAuth 2.1 auth, per-client isolation) and the Postgres application plane all work today and are covered by the suite. A real critique backend runs over a local verdict checkout (pnpm review <url>, see Getting real judgments). The main open gaps are design_recheck against a verdict backend, real evidence images in the panel, and moving the recheck index and unit ledger into the store. Full detail and the seam for each is in docs/roadmap.md.

Contributing

CONTRIBUTING.md covers setup, the test and lint commands, and the conventions that matter (the eyes-not-hands boundary, the contract-tested tool catalog, golden fixtures); the roadmap items are the best place to start, and it is worth opening an issue first if a change is large.

Security

Report vulnerabilities privately through GitHub's private vulnerability reporting on the repository's Security tab. SECURITY.md describes supported versions, what a reporter can expect, and what to check before running this against a network you care about.

License

MIT — see LICENSE.