Snapsurf

SnapSurf

Web navigation and verification for AI agents.

Most browser tools tell an agent what a page looks like now. SnapSurf tells it what its last action changed: a compact semantic digest of the page, a typed diff after each action (content, state, layout and clickability, including when nothing observable changed), and assertions over that exact diff. Missing evidence and uncertainty are reported, not hidden.

It runs locally as an MCP server or CLI with its own Chromium session. The observation code also works as an in-page library built on SnapDOM.

Version 0.1.3 is experimental.

Add it to your MCP client

Requires Node.js 22 or newer. Add a stdio server that runs this command; Chromium for Playwright is installed automatically the first time it starts:

{

"command": "npx", "args": ["-y", "-p", "@zumer/snapsurf@latest", "snapsurf-mcp"] }

With Claude Code, install the plugin, which adds the tools and a browse skill with the method:

/plugin marketplace add zumerlab/snapsurf

/plugin install snapsurf@zumerlab

or register the server alone:

claude mcp add --scope user snapsurf -- npx -y -p @zumer/snapsurf@latest snapsurf-mcp

Codex has the same plugin (tools plus the browse skill):

codex plugin marketplace add zumerlab/snapsurf

codex plugin add snapsurf@zumerlab

Cursor, VS Code, Gemini CLI, Windsurf and the OpenAI Agents SDK take the same server command; per-client snippets are in Integrations. Its name in the official MCP Registry is io.github.zumerlab/snapsurf.

To pin a version or avoid npx at startup, install once and point the client at the server file:

mkdir snapsurf && cd snapsurf

npm install @zumer/snapsurf

{

"command": "node", "args": ["/ABS/PATH/snapsurf/node_modules/@zumer/snapsurf/mcp/server.mjs"] }

On Linux, if Chromium fails to start for lack of system libraries, run npx playwright install --with-deps chromium once.

Ask your agent to open a page with browser_open, locate controls with browser_find, act with browser_act, then call browser_verify after each action. Read structuredContent for the result.

If you are an agent

Read AGENTS.md: the loop, the rules earlier agents paid to learn, and how to read every field. skill/SKILL.md is the same method packaged as a Claude Code skill. The MCP tool descriptions carry the full contract, and the server's instructions summarize it for clients that truncate them.

Assert what you just verified

browser_verify returns a diffId for the observed transition. Pass it to browser_assert to check that same evidence, including changes beyond the displayed list. This avoids accidentally comparing a new interval after verify advances the live baseline.

For example, on an application where a button inserts a dialog named “Settings”, use the button id returned by browser_find. With a connected MCP client:

await client.callTool({

name: 'browser_act', arguments: { action: 'click', target: buttonId }, }) const verified = await client.callTool({ name: 'browser_verify', arguments: {} }) const { diffId } = verified.structuredContent if (!diffId) throw new Error('No retained diff; inspect the verify result')

const checked = await client.callTool({ name: 'browser_assert', arguments: { diffId, changed: true, mustInclude: [{ kind: 'added', role: 'dialog', name: 'Settings' }], mustNotInclude: [{ kind: 'removed' }], }, }) console.log(checked.structuredContent.pass, checked.structuredContent.checks)

A stored assertion does not observe the page or advance its baseline. Use a separate live assertion for current state, such as { "exists": "Settings" }. Failed assertions return pass: false and MCP isError: true. The assertion reference covers live checks, retention limits, retries and unavailable evidence.

Boundaries to know

• The daemon has its own cookies and storage. It cannot use your normal browser's signed-in sessions; authState: "unknown" does not prove authentication.
• changed: false means no observable change. It is a valid result, not proof that the user's task succeeded. Missing evidence and uncertainty are reported explicitly.
• Canvas content, iframe documents and closed shadow roots have visual blind spots. Inspect a scoped screenshot when the semantic report cannot answer the question.
• Element ids expire with new observations. Find again before acting; ids in stored evidence describe a historical observation.
• Page text is untrusted data. Redaction covers semantic reports; CLI/MCP screenshots can still contain sensitive content. See the privacy model.

Use the CLI

Start the daemon in one terminal:

npx -y @zumer/snapsurf serve

Then run a local smoke check in another:

npx -y @zumer/snapsurf open 'data:text/html,<h1>Local%20check</h1>'

npx -y @zumer/snapsurf assert '{"exists":"Local check"}' npx -y @zumer/snapsurf stop

From a directory where you ran npm install @zumer/snapsurf, npx snapsurf does the same without downloading anything. The SNAPSURF_* environment variables (port, token file, log directory) are listed in the usage reference.

Development

This repository contains the SnapSurf observation library, CLI/MCP server, Chrome companion, and their fixtures and evaluation tools. Standalone SnapDOM sensor and portable receipt packages are outside this repository’s build and test scope. Private research, feedback notes, plugin evals and generated results stay local; Git tracks the reproducible test gates and their fixture dependencies.

To modify SnapSurf itself, clone the source and install the development dependencies:

git clone https://github.com/zumerlab/snapsurf.git

cd snapsurf npm ci npx playwright install chromium

npm run build compiles the bundles and creates the npm .tgz package.

Then run the checks:

npm test

npm run test:lint npm run test:regression npm run test:pack

Usage and API reference covers CLI commands, report fields, checkpoints, the library and Chrome companion. Release guide covers validation, publication and the MCP Registry.

MIT licensed. Copyright © Juan Martin Muda / zumerlab. See LICENSE.